linux/drivers/md/dm-cache-policy-smq.c

1762 lines
40 KiB
C
Raw Normal View History

dm cache: add stochastic-multi-queue (smq) policy The stochastic-multi-queue (smq) policy addresses some of the problems with the current multiqueue (mq) policy. Memory usage ------------ The mq policy uses a lot of memory; 88 bytes per cache block on a 64 bit machine. SMQ uses 28bit indexes to implement it's data structures rather than pointers. It avoids storing an explicit hit count for each block. It has a 'hotspot' queue rather than a pre cache which uses a quarter of the entries (each hotspot block covers a larger area than a single cache block). All these mean smq uses ~25bytes per cache block. Still a lot of memory, but a substantial improvement nontheless. Level balancing --------------- MQ places entries in different levels of the multiqueue structures based on their hit count (~ln(hit count)). This means the bottom levels generally have the most entries, and the top ones have very few. Having unbalanced levels like this reduces the efficacy of the multiqueue. SMQ does not maintain a hit count, instead it swaps hit entries with the least recently used entry from the level above. The over all ordering being a side effect of this stochastic process. With this scheme we can decide how many entries occupy each multiqueue level, resulting in better promotion/demotion decisions. Adaptability ------------ The MQ policy maintains a hit count for each cache block. For a different block to get promoted to the cache it's hit count has to exceed the lowest currently in the cache. This means it can take a long time for the cache to adapt between varying IO patterns. Periodically degrading the hit counts could help with this, but I haven't found a nice general solution. SMQ doesn't maintain hit counts, so a lot of this problem just goes away. In addition it tracks performance of the hotspot queue, which is used to decide which blocks to promote. If the hotspot queue is performing badly then it starts moving entries more quickly between levels. This lets it adapt to new IO patterns very quickly. Performance ----------- In my tests SMQ shows substantially better performance than MQ. Once this matures a bit more I'm sure it'll become the default policy. Signed-off-by: Joe Thornber <ejt@redhat.com> Signed-off-by: Mike Snitzer <snitzer@redhat.com>
2015-05-15 14:33:34 +00:00
/*
* Copyright (C) 2015 Red Hat. All rights reserved.
*
* This file is released under the GPL.
*/
#include "dm-cache-policy.h"
#include "dm-cache-policy-internal.h"
#include "dm.h"
#include <linux/hash.h>
#include <linux/jiffies.h>
#include <linux/module.h>
#include <linux/mutex.h>
#include <linux/vmalloc.h>
#include <linux/math64.h>
#define DM_MSG_PREFIX "cache-policy-smq"
/*----------------------------------------------------------------*/
/*
* Safe division functions that return zero on divide by zero.
*/
static unsigned safe_div(unsigned n, unsigned d)
{
return d ? n / d : 0u;
}
static unsigned safe_mod(unsigned n, unsigned d)
{
return d ? n % d : 0u;
}
/*----------------------------------------------------------------*/
struct entry {
unsigned hash_next:28;
unsigned prev:28;
unsigned next:28;
unsigned level:7;
bool dirty:1;
bool allocated:1;
bool sentinel:1;
dm_oblock_t oblock;
};
/*----------------------------------------------------------------*/
#define INDEXER_NULL ((1u << 28u) - 1u)
/*
* An entry_space manages a set of entries that we use for the queues.
* The clean and dirty queues share entries, so this object is separate
* from the queue itself.
*/
struct entry_space {
struct entry *begin;
struct entry *end;
};
static int space_init(struct entry_space *es, unsigned nr_entries)
{
if (!nr_entries) {
es->begin = es->end = NULL;
return 0;
}
es->begin = vzalloc(sizeof(struct entry) * nr_entries);
if (!es->begin)
return -ENOMEM;
es->end = es->begin + nr_entries;
return 0;
}
static void space_exit(struct entry_space *es)
{
vfree(es->begin);
}
static struct entry *__get_entry(struct entry_space *es, unsigned block)
{
struct entry *e;
e = es->begin + block;
BUG_ON(e >= es->end);
return e;
}
static unsigned to_index(struct entry_space *es, struct entry *e)
{
BUG_ON(e < es->begin || e >= es->end);
return e - es->begin;
}
static struct entry *to_entry(struct entry_space *es, unsigned block)
{
if (block == INDEXER_NULL)
return NULL;
return __get_entry(es, block);
}
/*----------------------------------------------------------------*/
struct ilist {
unsigned nr_elts; /* excluding sentinel entries */
unsigned head, tail;
};
static void l_init(struct ilist *l)
{
l->nr_elts = 0;
l->head = l->tail = INDEXER_NULL;
}
static struct entry *l_head(struct entry_space *es, struct ilist *l)
{
return to_entry(es, l->head);
}
static struct entry *l_tail(struct entry_space *es, struct ilist *l)
{
return to_entry(es, l->tail);
}
static struct entry *l_next(struct entry_space *es, struct entry *e)
{
return to_entry(es, e->next);
}
static struct entry *l_prev(struct entry_space *es, struct entry *e)
{
return to_entry(es, e->prev);
}
static bool l_empty(struct ilist *l)
{
return l->head == INDEXER_NULL;
}
static void l_add_head(struct entry_space *es, struct ilist *l, struct entry *e)
{
struct entry *head = l_head(es, l);
e->next = l->head;
e->prev = INDEXER_NULL;
if (head)
head->prev = l->head = to_index(es, e);
else
l->head = l->tail = to_index(es, e);
if (!e->sentinel)
l->nr_elts++;
}
static void l_add_tail(struct entry_space *es, struct ilist *l, struct entry *e)
{
struct entry *tail = l_tail(es, l);
e->next = INDEXER_NULL;
e->prev = l->tail;
if (tail)
tail->next = l->tail = to_index(es, e);
else
l->head = l->tail = to_index(es, e);
if (!e->sentinel)
l->nr_elts++;
}
static void l_add_before(struct entry_space *es, struct ilist *l,
struct entry *old, struct entry *e)
{
struct entry *prev = l_prev(es, old);
if (!prev)
l_add_head(es, l, e);
else {
e->prev = old->prev;
e->next = to_index(es, old);
prev->next = old->prev = to_index(es, e);
if (!e->sentinel)
l->nr_elts++;
}
}
static void l_del(struct entry_space *es, struct ilist *l, struct entry *e)
{
struct entry *prev = l_prev(es, e);
struct entry *next = l_next(es, e);
if (prev)
prev->next = e->next;
else
l->head = e->next;
if (next)
next->prev = e->prev;
else
l->tail = e->prev;
if (!e->sentinel)
l->nr_elts--;
}
static struct entry *l_pop_tail(struct entry_space *es, struct ilist *l)
{
struct entry *e;
for (e = l_tail(es, l); e; e = l_prev(es, e))
if (!e->sentinel) {
l_del(es, l, e);
return e;
}
return NULL;
}
/*----------------------------------------------------------------*/
/*
* The stochastic-multi-queue is a set of lru lists stacked into levels.
* Entries are moved up levels when they are used, which loosely orders the
* most accessed entries in the top levels and least in the bottom. This
* structure is *much* better than a single lru list.
*/
#define MAX_LEVELS 64u
struct queue {
struct entry_space *es;
unsigned nr_elts;
unsigned nr_levels;
struct ilist qs[MAX_LEVELS];
/*
* We maintain a count of the number of entries we would like in each
* level.
*/
unsigned last_target_nr_elts;
unsigned nr_top_levels;
unsigned nr_in_top_levels;
unsigned target_count[MAX_LEVELS];
};
static void q_init(struct queue *q, struct entry_space *es, unsigned nr_levels)
{
unsigned i;
q->es = es;
q->nr_elts = 0;
q->nr_levels = nr_levels;
for (i = 0; i < q->nr_levels; i++) {
l_init(q->qs + i);
q->target_count[i] = 0u;
}
q->last_target_nr_elts = 0u;
q->nr_top_levels = 0u;
q->nr_in_top_levels = 0u;
}
static unsigned q_size(struct queue *q)
{
return q->nr_elts;
}
/*
* Insert an entry to the back of the given level.
*/
static void q_push(struct queue *q, struct entry *e)
{
if (!e->sentinel)
q->nr_elts++;
l_add_tail(q->es, q->qs + e->level, e);
}
static void q_push_before(struct queue *q, struct entry *old, struct entry *e)
{
if (!e->sentinel)
q->nr_elts++;
l_add_before(q->es, q->qs + e->level, old, e);
}
static void q_del(struct queue *q, struct entry *e)
{
l_del(q->es, q->qs + e->level, e);
if (!e->sentinel)
q->nr_elts--;
}
/*
* Return the oldest entry of the lowest populated level.
*/
static struct entry *q_peek(struct queue *q, unsigned max_level, bool can_cross_sentinel)
{
unsigned level;
struct entry *e;
max_level = min(max_level, q->nr_levels);
for (level = 0; level < max_level; level++)
for (e = l_head(q->es, q->qs + level); e; e = l_next(q->es, e)) {
if (e->sentinel) {
if (can_cross_sentinel)
continue;
else
break;
}
return e;
}
return NULL;
}
static struct entry *q_pop(struct queue *q)
{
struct entry *e = q_peek(q, q->nr_levels, true);
if (e)
q_del(q, e);
return e;
}
/*
* Pops an entry from a level that is not past a sentinel.
*/
static struct entry *q_pop_old(struct queue *q, unsigned max_level)
{
struct entry *e = q_peek(q, max_level, false);
if (e)
q_del(q, e);
return e;
}
/*
* This function assumes there is a non-sentinel entry to pop. It's only
* used by redistribute, so we know this is true. It also doesn't adjust
* the q->nr_elts count.
*/
static struct entry *__redist_pop_from(struct queue *q, unsigned level)
{
struct entry *e;
for (; level < q->nr_levels; level++)
for (e = l_head(q->es, q->qs + level); e; e = l_next(q->es, e))
if (!e->sentinel) {
l_del(q->es, q->qs + e->level, e);
return e;
}
return NULL;
}
static void q_set_targets_subrange_(struct queue *q, unsigned nr_elts, unsigned lbegin, unsigned lend)
{
unsigned level, nr_levels, entries_per_level, remainder;
BUG_ON(lbegin > lend);
BUG_ON(lend > q->nr_levels);
nr_levels = lend - lbegin;
entries_per_level = safe_div(nr_elts, nr_levels);
remainder = safe_mod(nr_elts, nr_levels);
for (level = lbegin; level < lend; level++)
q->target_count[level] =
(level < (lbegin + remainder)) ? entries_per_level + 1u : entries_per_level;
}
/*
* Typically we have fewer elements in the top few levels which allows us
* to adjust the promote threshold nicely.
*/
static void q_set_targets(struct queue *q)
{
if (q->last_target_nr_elts == q->nr_elts)
return;
q->last_target_nr_elts = q->nr_elts;
if (q->nr_top_levels > q->nr_levels)
q_set_targets_subrange_(q, q->nr_elts, 0, q->nr_levels);
else {
q_set_targets_subrange_(q, q->nr_in_top_levels,
q->nr_levels - q->nr_top_levels, q->nr_levels);
if (q->nr_in_top_levels < q->nr_elts)
q_set_targets_subrange_(q, q->nr_elts - q->nr_in_top_levels,
0, q->nr_levels - q->nr_top_levels);
else
q_set_targets_subrange_(q, 0, 0, q->nr_levels - q->nr_top_levels);
}
}
static void q_redistribute(struct queue *q)
{
unsigned target, level;
struct ilist *l, *l_above;
struct entry *e;
q_set_targets(q);
for (level = 0u; level < q->nr_levels - 1u; level++) {
l = q->qs + level;
target = q->target_count[level];
/*
* Pull down some entries from the level above.
*/
while (l->nr_elts < target) {
e = __redist_pop_from(q, level + 1u);
if (!e) {
/* bug in nr_elts */
break;
}
e->level = level;
l_add_tail(q->es, l, e);
}
/*
* Push some entries up.
*/
l_above = q->qs + level + 1u;
while (l->nr_elts > target) {
e = l_pop_tail(q->es, l);
if (!e)
/* bug in nr_elts */
break;
e->level = level + 1u;
l_add_head(q->es, l_above, e);
}
}
}
static void q_requeue_before(struct queue *q, struct entry *dest, struct entry *e, unsigned extra_levels)
{
struct entry *de;
unsigned new_level;
q_del(q, e);
if (extra_levels && (e->level < q->nr_levels - 1u)) {
new_level = min(q->nr_levels - 1u, e->level + extra_levels);
for (de = l_head(q->es, q->qs + new_level); de; de = l_next(q->es, de)) {
if (de->sentinel)
continue;
q_del(q, de);
de->level = e->level;
if (dest)
q_push_before(q, dest, de);
else
q_push(q, de);
break;
}
e->level = new_level;
}
q_push(q, e);
}
static void q_requeue(struct queue *q, struct entry *e, unsigned extra_levels)
{
q_requeue_before(q, NULL, e, extra_levels);
}
/*----------------------------------------------------------------*/
#define FP_SHIFT 8
#define SIXTEENTH (1u << (FP_SHIFT - 4u))
#define EIGHTH (1u << (FP_SHIFT - 3u))
struct stats {
unsigned hit_threshold;
unsigned hits;
unsigned misses;
};
enum performance {
Q_POOR,
Q_FAIR,
Q_WELL
};
static void stats_init(struct stats *s, unsigned nr_levels)
{
s->hit_threshold = (nr_levels * 3u) / 4u;
s->hits = 0u;
s->misses = 0u;
}
static void stats_reset(struct stats *s)
{
s->hits = s->misses = 0u;
}
static void stats_level_accessed(struct stats *s, unsigned level)
{
if (level >= s->hit_threshold)
s->hits++;
else
s->misses++;
}
static void stats_miss(struct stats *s)
{
s->misses++;
}
/*
* There are times when we don't have any confidence in the hotspot queue.
* Such as when a fresh cache is created and the blocks have been spread
* out across the levels, or if an io load changes. We detect this by
* seeing how often a lookup is in the top levels of the hotspot queue.
*/
static enum performance stats_assess(struct stats *s)
{
unsigned confidence = safe_div(s->hits << FP_SHIFT, s->hits + s->misses);
if (confidence < SIXTEENTH)
return Q_POOR;
else if (confidence < EIGHTH)
return Q_FAIR;
else
return Q_WELL;
}
/*----------------------------------------------------------------*/
struct hash_table {
struct entry_space *es;
unsigned long long hash_bits;
unsigned *buckets;
};
/*
* All cache entries are stored in a chained hash table. To save space we
* use indexing again, and only store indexes to the next entry.
*/
static int h_init(struct hash_table *ht, struct entry_space *es, unsigned nr_entries)
{
unsigned i, nr_buckets;
ht->es = es;
nr_buckets = roundup_pow_of_two(max(nr_entries / 4u, 16u));
ht->hash_bits = ffs(nr_buckets) - 1;
ht->buckets = vmalloc(sizeof(*ht->buckets) * nr_buckets);
if (!ht->buckets)
return -ENOMEM;
for (i = 0; i < nr_buckets; i++)
ht->buckets[i] = INDEXER_NULL;
return 0;
}
static void h_exit(struct hash_table *ht)
{
vfree(ht->buckets);
}
static struct entry *h_head(struct hash_table *ht, unsigned bucket)
{
return to_entry(ht->es, ht->buckets[bucket]);
}
static struct entry *h_next(struct hash_table *ht, struct entry *e)
{
return to_entry(ht->es, e->hash_next);
}
static void __h_insert(struct hash_table *ht, unsigned bucket, struct entry *e)
{
e->hash_next = ht->buckets[bucket];
ht->buckets[bucket] = to_index(ht->es, e);
}
static void h_insert(struct hash_table *ht, struct entry *e)
{
unsigned h = hash_64(from_oblock(e->oblock), ht->hash_bits);
__h_insert(ht, h, e);
}
static struct entry *__h_lookup(struct hash_table *ht, unsigned h, dm_oblock_t oblock,
struct entry **prev)
{
struct entry *e;
*prev = NULL;
for (e = h_head(ht, h); e; e = h_next(ht, e)) {
if (e->oblock == oblock)
return e;
*prev = e;
}
return NULL;
}
static void __h_unlink(struct hash_table *ht, unsigned h,
struct entry *e, struct entry *prev)
{
if (prev)
prev->hash_next = e->hash_next;
else
ht->buckets[h] = e->hash_next;
}
/*
* Also moves each entry to the front of the bucket.
*/
static struct entry *h_lookup(struct hash_table *ht, dm_oblock_t oblock)
{
struct entry *e, *prev;
unsigned h = hash_64(from_oblock(oblock), ht->hash_bits);
e = __h_lookup(ht, h, oblock, &prev);
if (e && prev) {
/*
* Move to the front because this entry is likely
* to be hit again.
*/
__h_unlink(ht, h, e, prev);
__h_insert(ht, h, e);
}
return e;
}
static void h_remove(struct hash_table *ht, struct entry *e)
{
unsigned h = hash_64(from_oblock(e->oblock), ht->hash_bits);
struct entry *prev;
/*
* The down side of using a singly linked list is we have to
* iterate the bucket to remove an item.
*/
e = __h_lookup(ht, h, e->oblock, &prev);
if (e)
__h_unlink(ht, h, e, prev);
}
/*----------------------------------------------------------------*/
struct entry_alloc {
struct entry_space *es;
unsigned begin;
unsigned nr_allocated;
struct ilist free;
};
static void init_allocator(struct entry_alloc *ea, struct entry_space *es,
unsigned begin, unsigned end)
{
unsigned i;
ea->es = es;
ea->nr_allocated = 0u;
ea->begin = begin;
l_init(&ea->free);
for (i = begin; i != end; i++)
l_add_tail(ea->es, &ea->free, __get_entry(ea->es, i));
}
static void init_entry(struct entry *e)
{
/*
* We can't memset because that would clear the hotspot and
* sentinel bits which remain constant.
*/
e->hash_next = INDEXER_NULL;
e->next = INDEXER_NULL;
e->prev = INDEXER_NULL;
e->level = 0u;
e->allocated = true;
}
static struct entry *alloc_entry(struct entry_alloc *ea)
{
struct entry *e;
if (l_empty(&ea->free))
return NULL;
e = l_pop_tail(ea->es, &ea->free);
init_entry(e);
ea->nr_allocated++;
return e;
}
/*
* This assumes the cblock hasn't already been allocated.
*/
static struct entry *alloc_particular_entry(struct entry_alloc *ea, unsigned i)
{
struct entry *e = __get_entry(ea->es, ea->begin + i);
BUG_ON(e->allocated);
l_del(ea->es, &ea->free, e);
init_entry(e);
ea->nr_allocated++;
return e;
}
static void free_entry(struct entry_alloc *ea, struct entry *e)
{
BUG_ON(!ea->nr_allocated);
BUG_ON(!e->allocated);
ea->nr_allocated--;
e->allocated = false;
l_add_tail(ea->es, &ea->free, e);
}
static bool allocator_empty(struct entry_alloc *ea)
{
return l_empty(&ea->free);
}
static unsigned get_index(struct entry_alloc *ea, struct entry *e)
{
return to_index(ea->es, e) - ea->begin;
}
static struct entry *get_entry(struct entry_alloc *ea, unsigned index)
{
return __get_entry(ea->es, ea->begin + index);
}
/*----------------------------------------------------------------*/
#define NR_HOTSPOT_LEVELS 64u
#define NR_CACHE_LEVELS 64u
#define WRITEBACK_PERIOD (10 * HZ)
#define DEMOTE_PERIOD (60 * HZ)
#define HOTSPOT_UPDATE_PERIOD (HZ)
#define CACHE_UPDATE_PERIOD (10u * HZ)
struct smq_policy {
struct dm_cache_policy policy;
/* protects everything */
spinlock_t lock;
dm cache: add stochastic-multi-queue (smq) policy The stochastic-multi-queue (smq) policy addresses some of the problems with the current multiqueue (mq) policy. Memory usage ------------ The mq policy uses a lot of memory; 88 bytes per cache block on a 64 bit machine. SMQ uses 28bit indexes to implement it's data structures rather than pointers. It avoids storing an explicit hit count for each block. It has a 'hotspot' queue rather than a pre cache which uses a quarter of the entries (each hotspot block covers a larger area than a single cache block). All these mean smq uses ~25bytes per cache block. Still a lot of memory, but a substantial improvement nontheless. Level balancing --------------- MQ places entries in different levels of the multiqueue structures based on their hit count (~ln(hit count)). This means the bottom levels generally have the most entries, and the top ones have very few. Having unbalanced levels like this reduces the efficacy of the multiqueue. SMQ does not maintain a hit count, instead it swaps hit entries with the least recently used entry from the level above. The over all ordering being a side effect of this stochastic process. With this scheme we can decide how many entries occupy each multiqueue level, resulting in better promotion/demotion decisions. Adaptability ------------ The MQ policy maintains a hit count for each cache block. For a different block to get promoted to the cache it's hit count has to exceed the lowest currently in the cache. This means it can take a long time for the cache to adapt between varying IO patterns. Periodically degrading the hit counts could help with this, but I haven't found a nice general solution. SMQ doesn't maintain hit counts, so a lot of this problem just goes away. In addition it tracks performance of the hotspot queue, which is used to decide which blocks to promote. If the hotspot queue is performing badly then it starts moving entries more quickly between levels. This lets it adapt to new IO patterns very quickly. Performance ----------- In my tests SMQ shows substantially better performance than MQ. Once this matures a bit more I'm sure it'll become the default policy. Signed-off-by: Joe Thornber <ejt@redhat.com> Signed-off-by: Mike Snitzer <snitzer@redhat.com>
2015-05-15 14:33:34 +00:00
dm_cblock_t cache_size;
sector_t cache_block_size;
sector_t hotspot_block_size;
unsigned nr_hotspot_blocks;
unsigned cache_blocks_per_hotspot_block;
unsigned hotspot_level_jump;
struct entry_space es;
struct entry_alloc writeback_sentinel_alloc;
struct entry_alloc demote_sentinel_alloc;
struct entry_alloc hotspot_alloc;
struct entry_alloc cache_alloc;
unsigned long *hotspot_hit_bits;
unsigned long *cache_hit_bits;
/*
* We maintain three queues of entries. The cache proper,
* consisting of a clean and dirty queue, containing the currently
* active mappings. The hotspot queue uses a larger block size to
* track blocks that are being hit frequently and potential
* candidates for promotion to the cache.
*/
struct queue hotspot;
struct queue clean;
struct queue dirty;
struct stats hotspot_stats;
struct stats cache_stats;
/*
* Keeps track of time, incremented by the core. We use this to
* avoid attributing multiple hits within the same tick.
*/
unsigned tick;
/*
* The hash tables allows us to quickly find an entry by origin
* block.
*/
struct hash_table table;
struct hash_table hotspot_table;
bool current_writeback_sentinels;
unsigned long next_writeback_period;
bool current_demote_sentinels;
unsigned long next_demote_period;
unsigned write_promote_level;
unsigned read_promote_level;
unsigned long next_hotspot_period;
unsigned long next_cache_period;
};
/*----------------------------------------------------------------*/
static struct entry *get_sentinel(struct entry_alloc *ea, unsigned level, bool which)
{
return get_entry(ea, which ? level : NR_CACHE_LEVELS + level);
}
static struct entry *writeback_sentinel(struct smq_policy *mq, unsigned level)
{
return get_sentinel(&mq->writeback_sentinel_alloc, level, mq->current_writeback_sentinels);
}
static struct entry *demote_sentinel(struct smq_policy *mq, unsigned level)
{
return get_sentinel(&mq->demote_sentinel_alloc, level, mq->current_demote_sentinels);
}
static void __update_writeback_sentinels(struct smq_policy *mq)
{
unsigned level;
struct queue *q = &mq->dirty;
struct entry *sentinel;
for (level = 0; level < q->nr_levels; level++) {
sentinel = writeback_sentinel(mq, level);
q_del(q, sentinel);
q_push(q, sentinel);
}
}
static void __update_demote_sentinels(struct smq_policy *mq)
{
unsigned level;
struct queue *q = &mq->clean;
struct entry *sentinel;
for (level = 0; level < q->nr_levels; level++) {
sentinel = demote_sentinel(mq, level);
q_del(q, sentinel);
q_push(q, sentinel);
}
}
static void update_sentinels(struct smq_policy *mq)
{
if (time_after(jiffies, mq->next_writeback_period)) {
__update_writeback_sentinels(mq);
mq->next_writeback_period = jiffies + WRITEBACK_PERIOD;
mq->current_writeback_sentinels = !mq->current_writeback_sentinels;
}
if (time_after(jiffies, mq->next_demote_period)) {
__update_demote_sentinels(mq);
mq->next_demote_period = jiffies + DEMOTE_PERIOD;
mq->current_demote_sentinels = !mq->current_demote_sentinels;
}
}
static void __sentinels_init(struct smq_policy *mq)
{
unsigned level;
struct entry *sentinel;
for (level = 0; level < NR_CACHE_LEVELS; level++) {
sentinel = writeback_sentinel(mq, level);
sentinel->level = level;
q_push(&mq->dirty, sentinel);
sentinel = demote_sentinel(mq, level);
sentinel->level = level;
q_push(&mq->clean, sentinel);
}
}
static void sentinels_init(struct smq_policy *mq)
{
mq->next_writeback_period = jiffies + WRITEBACK_PERIOD;
mq->next_demote_period = jiffies + DEMOTE_PERIOD;
mq->current_writeback_sentinels = false;
mq->current_demote_sentinels = false;
__sentinels_init(mq);
mq->current_writeback_sentinels = !mq->current_writeback_sentinels;
mq->current_demote_sentinels = !mq->current_demote_sentinels;
__sentinels_init(mq);
}
/*----------------------------------------------------------------*/
/*
* These methods tie together the dirty queue, clean queue and hash table.
*/
static void push_new(struct smq_policy *mq, struct entry *e)
{
struct queue *q = e->dirty ? &mq->dirty : &mq->clean;
h_insert(&mq->table, e);
q_push(q, e);
}
static void push(struct smq_policy *mq, struct entry *e)
{
struct entry *sentinel;
h_insert(&mq->table, e);
/*
* Punch this into the queue just in front of the sentinel, to
* ensure it's cleaned straight away.
*/
if (e->dirty) {
sentinel = writeback_sentinel(mq, e->level);
q_push_before(&mq->dirty, sentinel, e);
} else {
sentinel = demote_sentinel(mq, e->level);
q_push_before(&mq->clean, sentinel, e);
}
}
/*
* Removes an entry from cache. Removes from the hash table.
*/
static void __del(struct smq_policy *mq, struct queue *q, struct entry *e)
{
q_del(q, e);
h_remove(&mq->table, e);
}
static void del(struct smq_policy *mq, struct entry *e)
{
__del(mq, e->dirty ? &mq->dirty : &mq->clean, e);
}
static struct entry *pop_old(struct smq_policy *mq, struct queue *q, unsigned max_level)
{
struct entry *e = q_pop_old(q, max_level);
if (e)
h_remove(&mq->table, e);
return e;
}
static dm_cblock_t infer_cblock(struct smq_policy *mq, struct entry *e)
{
return to_cblock(get_index(&mq->cache_alloc, e));
}
static void requeue(struct smq_policy *mq, struct entry *e)
{
struct entry *sentinel;
if (!test_and_set_bit(from_cblock(infer_cblock(mq, e)), mq->cache_hit_bits)) {
if (e->dirty) {
sentinel = writeback_sentinel(mq, e->level);
q_requeue_before(&mq->dirty, sentinel, e, 1u);
} else {
sentinel = demote_sentinel(mq, e->level);
q_requeue_before(&mq->clean, sentinel, e, 1u);
}
}
}
static unsigned default_promote_level(struct smq_policy *mq)
{
/*
* The promote level depends on the current performance of the
* cache.
*
* If the cache is performing badly, then we can't afford
* to promote much without causing performance to drop below that
* of the origin device.
*
* If the cache is performing well, then we don't need to promote
* much. If it isn't broken, don't fix it.
*
* If the cache is middling then we promote more.
*
* This scheme reminds me of a graph of entropy vs probability of a
* binary variable.
*/
static unsigned table[] = {1, 1, 1, 2, 4, 6, 7, 8, 7, 6, 4, 4, 3, 3, 2, 2, 1};
unsigned hits = mq->cache_stats.hits;
unsigned misses = mq->cache_stats.misses;
unsigned index = safe_div(hits << 4u, hits + misses);
return table[index];
}
static void update_promote_levels(struct smq_policy *mq)
{
/*
* If there are unused cache entries then we want to be really
* eager to promote.
*/
unsigned threshold_level = allocator_empty(&mq->cache_alloc) ?
default_promote_level(mq) : (NR_HOTSPOT_LEVELS / 2u);
/*
* If the hotspot queue is performing badly then we have little
* confidence that we know which blocks to promote. So we cut down
* the amount of promotions.
*/
switch (stats_assess(&mq->hotspot_stats)) {
case Q_POOR:
threshold_level /= 4u;
break;
case Q_FAIR:
threshold_level /= 2u;
break;
case Q_WELL:
break;
}
mq->read_promote_level = NR_HOTSPOT_LEVELS - threshold_level;
mq->write_promote_level = (NR_HOTSPOT_LEVELS - threshold_level) + 2u;
}
/*
* If the hotspot queue is performing badly, then we try and move entries
* around more quickly.
*/
static void update_level_jump(struct smq_policy *mq)
{
switch (stats_assess(&mq->hotspot_stats)) {
case Q_POOR:
mq->hotspot_level_jump = 4u;
break;
case Q_FAIR:
mq->hotspot_level_jump = 2u;
break;
case Q_WELL:
mq->hotspot_level_jump = 1u;
break;
}
}
static void end_hotspot_period(struct smq_policy *mq)
{
clear_bitset(mq->hotspot_hit_bits, mq->nr_hotspot_blocks);
update_promote_levels(mq);
if (time_after(jiffies, mq->next_hotspot_period)) {
update_level_jump(mq);
q_redistribute(&mq->hotspot);
stats_reset(&mq->hotspot_stats);
mq->next_hotspot_period = jiffies + HOTSPOT_UPDATE_PERIOD;
}
}
static void end_cache_period(struct smq_policy *mq)
{
if (time_after(jiffies, mq->next_cache_period)) {
clear_bitset(mq->cache_hit_bits, from_cblock(mq->cache_size));
q_redistribute(&mq->dirty);
q_redistribute(&mq->clean);
stats_reset(&mq->cache_stats);
mq->next_cache_period = jiffies + CACHE_UPDATE_PERIOD;
}
}
static int demote_cblock(struct smq_policy *mq,
struct policy_locker *locker,
dm_oblock_t *oblock)
{
struct entry *demoted = q_peek(&mq->clean, mq->clean.nr_levels, false);
if (!demoted)
/*
* We could get a block from mq->dirty, but that
* would add extra latency to the triggering bio as it
* waits for the writeback. Better to not promote this
* time and hope there's a clean block next time this block
* is hit.
*/
return -ENOSPC;
if (locker->fn(locker, demoted->oblock))
/*
* We couldn't lock this block.
*/
return -EBUSY;
del(mq, demoted);
*oblock = demoted->oblock;
free_entry(&mq->cache_alloc, demoted);
return 0;
}
enum promote_result {
PROMOTE_NOT,
PROMOTE_TEMPORARY,
PROMOTE_PERMANENT
};
/*
* Converts a boolean into a promote result.
*/
static enum promote_result maybe_promote(bool promote)
{
return promote ? PROMOTE_PERMANENT : PROMOTE_NOT;
}
static enum promote_result should_promote(struct smq_policy *mq, struct entry *hs_e, struct bio *bio,
bool fast_promote)
{
if (bio_data_dir(bio) == WRITE) {
if (!allocator_empty(&mq->cache_alloc) && fast_promote)
return PROMOTE_TEMPORARY;
else
return maybe_promote(hs_e->level >= mq->write_promote_level);
} else
return maybe_promote(hs_e->level >= mq->read_promote_level);
}
static void insert_in_cache(struct smq_policy *mq, dm_oblock_t oblock,
struct policy_locker *locker,
struct policy_result *result, enum promote_result pr)
{
int r;
struct entry *e;
if (allocator_empty(&mq->cache_alloc)) {
result->op = POLICY_REPLACE;
r = demote_cblock(mq, locker, &result->old_oblock);
if (r) {
result->op = POLICY_MISS;
return;
}
} else
result->op = POLICY_NEW;
e = alloc_entry(&mq->cache_alloc);
BUG_ON(!e);
e->oblock = oblock;
if (pr == PROMOTE_TEMPORARY)
push(mq, e);
else
push_new(mq, e);
result->cblock = infer_cblock(mq, e);
}
static dm_oblock_t to_hblock(struct smq_policy *mq, dm_oblock_t b)
{
sector_t r = from_oblock(b);
(void) sector_div(r, mq->cache_blocks_per_hotspot_block);
return to_oblock(r);
}
static struct entry *update_hotspot_queue(struct smq_policy *mq, dm_oblock_t b, struct bio *bio)
{
unsigned hi;
dm_oblock_t hb = to_hblock(mq, b);
struct entry *e = h_lookup(&mq->hotspot_table, hb);
if (e) {
stats_level_accessed(&mq->hotspot_stats, e->level);
hi = get_index(&mq->hotspot_alloc, e);
q_requeue(&mq->hotspot, e,
test_and_set_bit(hi, mq->hotspot_hit_bits) ?
0u : mq->hotspot_level_jump);
} else {
stats_miss(&mq->hotspot_stats);
e = alloc_entry(&mq->hotspot_alloc);
if (!e) {
e = q_pop(&mq->hotspot);
if (e) {
h_remove(&mq->hotspot_table, e);
hi = get_index(&mq->hotspot_alloc, e);
clear_bit(hi, mq->hotspot_hit_bits);
}
}
if (e) {
e->oblock = hb;
q_push(&mq->hotspot, e);
h_insert(&mq->hotspot_table, e);
}
}
return e;
}
/*
* Looks the oblock up in the hash table, then decides whether to put in
* pre_cache, or cache etc.
*/
static int map(struct smq_policy *mq, struct bio *bio, dm_oblock_t oblock,
bool can_migrate, bool fast_promote,
struct policy_locker *locker, struct policy_result *result)
{
struct entry *e, *hs_e;
enum promote_result pr;
hs_e = update_hotspot_queue(mq, oblock, bio);
e = h_lookup(&mq->table, oblock);
if (e) {
stats_level_accessed(&mq->cache_stats, e->level);
requeue(mq, e);
result->op = POLICY_HIT;
result->cblock = infer_cblock(mq, e);
} else {
stats_miss(&mq->cache_stats);
pr = should_promote(mq, hs_e, bio, fast_promote);
if (pr == PROMOTE_NOT)
result->op = POLICY_MISS;
else {
if (!can_migrate) {
result->op = POLICY_MISS;
return -EWOULDBLOCK;
}
insert_in_cache(mq, oblock, locker, result, pr);
}
}
return 0;
}
/*----------------------------------------------------------------*/
/*
* Public interface, via the policy struct. See dm-cache-policy.h for a
* description of these.
*/
static struct smq_policy *to_smq_policy(struct dm_cache_policy *p)
{
return container_of(p, struct smq_policy, policy);
}
static void smq_destroy(struct dm_cache_policy *p)
{
struct smq_policy *mq = to_smq_policy(p);
h_exit(&mq->hotspot_table);
h_exit(&mq->table);
free_bitset(mq->hotspot_hit_bits);
free_bitset(mq->cache_hit_bits);
space_exit(&mq->es);
kfree(mq);
}
static int smq_map(struct dm_cache_policy *p, dm_oblock_t oblock,
bool can_block, bool can_migrate, bool fast_promote,
struct bio *bio, struct policy_locker *locker,
struct policy_result *result)
{
int r;
unsigned long flags;
dm cache: add stochastic-multi-queue (smq) policy The stochastic-multi-queue (smq) policy addresses some of the problems with the current multiqueue (mq) policy. Memory usage ------------ The mq policy uses a lot of memory; 88 bytes per cache block on a 64 bit machine. SMQ uses 28bit indexes to implement it's data structures rather than pointers. It avoids storing an explicit hit count for each block. It has a 'hotspot' queue rather than a pre cache which uses a quarter of the entries (each hotspot block covers a larger area than a single cache block). All these mean smq uses ~25bytes per cache block. Still a lot of memory, but a substantial improvement nontheless. Level balancing --------------- MQ places entries in different levels of the multiqueue structures based on their hit count (~ln(hit count)). This means the bottom levels generally have the most entries, and the top ones have very few. Having unbalanced levels like this reduces the efficacy of the multiqueue. SMQ does not maintain a hit count, instead it swaps hit entries with the least recently used entry from the level above. The over all ordering being a side effect of this stochastic process. With this scheme we can decide how many entries occupy each multiqueue level, resulting in better promotion/demotion decisions. Adaptability ------------ The MQ policy maintains a hit count for each cache block. For a different block to get promoted to the cache it's hit count has to exceed the lowest currently in the cache. This means it can take a long time for the cache to adapt between varying IO patterns. Periodically degrading the hit counts could help with this, but I haven't found a nice general solution. SMQ doesn't maintain hit counts, so a lot of this problem just goes away. In addition it tracks performance of the hotspot queue, which is used to decide which blocks to promote. If the hotspot queue is performing badly then it starts moving entries more quickly between levels. This lets it adapt to new IO patterns very quickly. Performance ----------- In my tests SMQ shows substantially better performance than MQ. Once this matures a bit more I'm sure it'll become the default policy. Signed-off-by: Joe Thornber <ejt@redhat.com> Signed-off-by: Mike Snitzer <snitzer@redhat.com>
2015-05-15 14:33:34 +00:00
struct smq_policy *mq = to_smq_policy(p);
result->op = POLICY_MISS;
spin_lock_irqsave(&mq->lock, flags);
dm cache: add stochastic-multi-queue (smq) policy The stochastic-multi-queue (smq) policy addresses some of the problems with the current multiqueue (mq) policy. Memory usage ------------ The mq policy uses a lot of memory; 88 bytes per cache block on a 64 bit machine. SMQ uses 28bit indexes to implement it's data structures rather than pointers. It avoids storing an explicit hit count for each block. It has a 'hotspot' queue rather than a pre cache which uses a quarter of the entries (each hotspot block covers a larger area than a single cache block). All these mean smq uses ~25bytes per cache block. Still a lot of memory, but a substantial improvement nontheless. Level balancing --------------- MQ places entries in different levels of the multiqueue structures based on their hit count (~ln(hit count)). This means the bottom levels generally have the most entries, and the top ones have very few. Having unbalanced levels like this reduces the efficacy of the multiqueue. SMQ does not maintain a hit count, instead it swaps hit entries with the least recently used entry from the level above. The over all ordering being a side effect of this stochastic process. With this scheme we can decide how many entries occupy each multiqueue level, resulting in better promotion/demotion decisions. Adaptability ------------ The MQ policy maintains a hit count for each cache block. For a different block to get promoted to the cache it's hit count has to exceed the lowest currently in the cache. This means it can take a long time for the cache to adapt between varying IO patterns. Periodically degrading the hit counts could help with this, but I haven't found a nice general solution. SMQ doesn't maintain hit counts, so a lot of this problem just goes away. In addition it tracks performance of the hotspot queue, which is used to decide which blocks to promote. If the hotspot queue is performing badly then it starts moving entries more quickly between levels. This lets it adapt to new IO patterns very quickly. Performance ----------- In my tests SMQ shows substantially better performance than MQ. Once this matures a bit more I'm sure it'll become the default policy. Signed-off-by: Joe Thornber <ejt@redhat.com> Signed-off-by: Mike Snitzer <snitzer@redhat.com>
2015-05-15 14:33:34 +00:00
r = map(mq, bio, oblock, can_migrate, fast_promote, locker, result);
spin_unlock_irqrestore(&mq->lock, flags);
dm cache: add stochastic-multi-queue (smq) policy The stochastic-multi-queue (smq) policy addresses some of the problems with the current multiqueue (mq) policy. Memory usage ------------ The mq policy uses a lot of memory; 88 bytes per cache block on a 64 bit machine. SMQ uses 28bit indexes to implement it's data structures rather than pointers. It avoids storing an explicit hit count for each block. It has a 'hotspot' queue rather than a pre cache which uses a quarter of the entries (each hotspot block covers a larger area than a single cache block). All these mean smq uses ~25bytes per cache block. Still a lot of memory, but a substantial improvement nontheless. Level balancing --------------- MQ places entries in different levels of the multiqueue structures based on their hit count (~ln(hit count)). This means the bottom levels generally have the most entries, and the top ones have very few. Having unbalanced levels like this reduces the efficacy of the multiqueue. SMQ does not maintain a hit count, instead it swaps hit entries with the least recently used entry from the level above. The over all ordering being a side effect of this stochastic process. With this scheme we can decide how many entries occupy each multiqueue level, resulting in better promotion/demotion decisions. Adaptability ------------ The MQ policy maintains a hit count for each cache block. For a different block to get promoted to the cache it's hit count has to exceed the lowest currently in the cache. This means it can take a long time for the cache to adapt between varying IO patterns. Periodically degrading the hit counts could help with this, but I haven't found a nice general solution. SMQ doesn't maintain hit counts, so a lot of this problem just goes away. In addition it tracks performance of the hotspot queue, which is used to decide which blocks to promote. If the hotspot queue is performing badly then it starts moving entries more quickly between levels. This lets it adapt to new IO patterns very quickly. Performance ----------- In my tests SMQ shows substantially better performance than MQ. Once this matures a bit more I'm sure it'll become the default policy. Signed-off-by: Joe Thornber <ejt@redhat.com> Signed-off-by: Mike Snitzer <snitzer@redhat.com>
2015-05-15 14:33:34 +00:00
return r;
}
static int smq_lookup(struct dm_cache_policy *p, dm_oblock_t oblock, dm_cblock_t *cblock)
{
int r;
unsigned long flags;
dm cache: add stochastic-multi-queue (smq) policy The stochastic-multi-queue (smq) policy addresses some of the problems with the current multiqueue (mq) policy. Memory usage ------------ The mq policy uses a lot of memory; 88 bytes per cache block on a 64 bit machine. SMQ uses 28bit indexes to implement it's data structures rather than pointers. It avoids storing an explicit hit count for each block. It has a 'hotspot' queue rather than a pre cache which uses a quarter of the entries (each hotspot block covers a larger area than a single cache block). All these mean smq uses ~25bytes per cache block. Still a lot of memory, but a substantial improvement nontheless. Level balancing --------------- MQ places entries in different levels of the multiqueue structures based on their hit count (~ln(hit count)). This means the bottom levels generally have the most entries, and the top ones have very few. Having unbalanced levels like this reduces the efficacy of the multiqueue. SMQ does not maintain a hit count, instead it swaps hit entries with the least recently used entry from the level above. The over all ordering being a side effect of this stochastic process. With this scheme we can decide how many entries occupy each multiqueue level, resulting in better promotion/demotion decisions. Adaptability ------------ The MQ policy maintains a hit count for each cache block. For a different block to get promoted to the cache it's hit count has to exceed the lowest currently in the cache. This means it can take a long time for the cache to adapt between varying IO patterns. Periodically degrading the hit counts could help with this, but I haven't found a nice general solution. SMQ doesn't maintain hit counts, so a lot of this problem just goes away. In addition it tracks performance of the hotspot queue, which is used to decide which blocks to promote. If the hotspot queue is performing badly then it starts moving entries more quickly between levels. This lets it adapt to new IO patterns very quickly. Performance ----------- In my tests SMQ shows substantially better performance than MQ. Once this matures a bit more I'm sure it'll become the default policy. Signed-off-by: Joe Thornber <ejt@redhat.com> Signed-off-by: Mike Snitzer <snitzer@redhat.com>
2015-05-15 14:33:34 +00:00
struct smq_policy *mq = to_smq_policy(p);
struct entry *e;
spin_lock_irqsave(&mq->lock, flags);
dm cache: add stochastic-multi-queue (smq) policy The stochastic-multi-queue (smq) policy addresses some of the problems with the current multiqueue (mq) policy. Memory usage ------------ The mq policy uses a lot of memory; 88 bytes per cache block on a 64 bit machine. SMQ uses 28bit indexes to implement it's data structures rather than pointers. It avoids storing an explicit hit count for each block. It has a 'hotspot' queue rather than a pre cache which uses a quarter of the entries (each hotspot block covers a larger area than a single cache block). All these mean smq uses ~25bytes per cache block. Still a lot of memory, but a substantial improvement nontheless. Level balancing --------------- MQ places entries in different levels of the multiqueue structures based on their hit count (~ln(hit count)). This means the bottom levels generally have the most entries, and the top ones have very few. Having unbalanced levels like this reduces the efficacy of the multiqueue. SMQ does not maintain a hit count, instead it swaps hit entries with the least recently used entry from the level above. The over all ordering being a side effect of this stochastic process. With this scheme we can decide how many entries occupy each multiqueue level, resulting in better promotion/demotion decisions. Adaptability ------------ The MQ policy maintains a hit count for each cache block. For a different block to get promoted to the cache it's hit count has to exceed the lowest currently in the cache. This means it can take a long time for the cache to adapt between varying IO patterns. Periodically degrading the hit counts could help with this, but I haven't found a nice general solution. SMQ doesn't maintain hit counts, so a lot of this problem just goes away. In addition it tracks performance of the hotspot queue, which is used to decide which blocks to promote. If the hotspot queue is performing badly then it starts moving entries more quickly between levels. This lets it adapt to new IO patterns very quickly. Performance ----------- In my tests SMQ shows substantially better performance than MQ. Once this matures a bit more I'm sure it'll become the default policy. Signed-off-by: Joe Thornber <ejt@redhat.com> Signed-off-by: Mike Snitzer <snitzer@redhat.com>
2015-05-15 14:33:34 +00:00
e = h_lookup(&mq->table, oblock);
if (e) {
*cblock = infer_cblock(mq, e);
r = 0;
} else
r = -ENOENT;
spin_unlock_irqrestore(&mq->lock, flags);
dm cache: add stochastic-multi-queue (smq) policy The stochastic-multi-queue (smq) policy addresses some of the problems with the current multiqueue (mq) policy. Memory usage ------------ The mq policy uses a lot of memory; 88 bytes per cache block on a 64 bit machine. SMQ uses 28bit indexes to implement it's data structures rather than pointers. It avoids storing an explicit hit count for each block. It has a 'hotspot' queue rather than a pre cache which uses a quarter of the entries (each hotspot block covers a larger area than a single cache block). All these mean smq uses ~25bytes per cache block. Still a lot of memory, but a substantial improvement nontheless. Level balancing --------------- MQ places entries in different levels of the multiqueue structures based on their hit count (~ln(hit count)). This means the bottom levels generally have the most entries, and the top ones have very few. Having unbalanced levels like this reduces the efficacy of the multiqueue. SMQ does not maintain a hit count, instead it swaps hit entries with the least recently used entry from the level above. The over all ordering being a side effect of this stochastic process. With this scheme we can decide how many entries occupy each multiqueue level, resulting in better promotion/demotion decisions. Adaptability ------------ The MQ policy maintains a hit count for each cache block. For a different block to get promoted to the cache it's hit count has to exceed the lowest currently in the cache. This means it can take a long time for the cache to adapt between varying IO patterns. Periodically degrading the hit counts could help with this, but I haven't found a nice general solution. SMQ doesn't maintain hit counts, so a lot of this problem just goes away. In addition it tracks performance of the hotspot queue, which is used to decide which blocks to promote. If the hotspot queue is performing badly then it starts moving entries more quickly between levels. This lets it adapt to new IO patterns very quickly. Performance ----------- In my tests SMQ shows substantially better performance than MQ. Once this matures a bit more I'm sure it'll become the default policy. Signed-off-by: Joe Thornber <ejt@redhat.com> Signed-off-by: Mike Snitzer <snitzer@redhat.com>
2015-05-15 14:33:34 +00:00
return r;
}
static void __smq_set_clear_dirty(struct smq_policy *mq, dm_oblock_t oblock, bool set)
{
struct entry *e;
e = h_lookup(&mq->table, oblock);
BUG_ON(!e);
del(mq, e);
e->dirty = set;
push(mq, e);
}
static void smq_set_dirty(struct dm_cache_policy *p, dm_oblock_t oblock)
{
unsigned long flags;
dm cache: add stochastic-multi-queue (smq) policy The stochastic-multi-queue (smq) policy addresses some of the problems with the current multiqueue (mq) policy. Memory usage ------------ The mq policy uses a lot of memory; 88 bytes per cache block on a 64 bit machine. SMQ uses 28bit indexes to implement it's data structures rather than pointers. It avoids storing an explicit hit count for each block. It has a 'hotspot' queue rather than a pre cache which uses a quarter of the entries (each hotspot block covers a larger area than a single cache block). All these mean smq uses ~25bytes per cache block. Still a lot of memory, but a substantial improvement nontheless. Level balancing --------------- MQ places entries in different levels of the multiqueue structures based on their hit count (~ln(hit count)). This means the bottom levels generally have the most entries, and the top ones have very few. Having unbalanced levels like this reduces the efficacy of the multiqueue. SMQ does not maintain a hit count, instead it swaps hit entries with the least recently used entry from the level above. The over all ordering being a side effect of this stochastic process. With this scheme we can decide how many entries occupy each multiqueue level, resulting in better promotion/demotion decisions. Adaptability ------------ The MQ policy maintains a hit count for each cache block. For a different block to get promoted to the cache it's hit count has to exceed the lowest currently in the cache. This means it can take a long time for the cache to adapt between varying IO patterns. Periodically degrading the hit counts could help with this, but I haven't found a nice general solution. SMQ doesn't maintain hit counts, so a lot of this problem just goes away. In addition it tracks performance of the hotspot queue, which is used to decide which blocks to promote. If the hotspot queue is performing badly then it starts moving entries more quickly between levels. This lets it adapt to new IO patterns very quickly. Performance ----------- In my tests SMQ shows substantially better performance than MQ. Once this matures a bit more I'm sure it'll become the default policy. Signed-off-by: Joe Thornber <ejt@redhat.com> Signed-off-by: Mike Snitzer <snitzer@redhat.com>
2015-05-15 14:33:34 +00:00
struct smq_policy *mq = to_smq_policy(p);
spin_lock_irqsave(&mq->lock, flags);
dm cache: add stochastic-multi-queue (smq) policy The stochastic-multi-queue (smq) policy addresses some of the problems with the current multiqueue (mq) policy. Memory usage ------------ The mq policy uses a lot of memory; 88 bytes per cache block on a 64 bit machine. SMQ uses 28bit indexes to implement it's data structures rather than pointers. It avoids storing an explicit hit count for each block. It has a 'hotspot' queue rather than a pre cache which uses a quarter of the entries (each hotspot block covers a larger area than a single cache block). All these mean smq uses ~25bytes per cache block. Still a lot of memory, but a substantial improvement nontheless. Level balancing --------------- MQ places entries in different levels of the multiqueue structures based on their hit count (~ln(hit count)). This means the bottom levels generally have the most entries, and the top ones have very few. Having unbalanced levels like this reduces the efficacy of the multiqueue. SMQ does not maintain a hit count, instead it swaps hit entries with the least recently used entry from the level above. The over all ordering being a side effect of this stochastic process. With this scheme we can decide how many entries occupy each multiqueue level, resulting in better promotion/demotion decisions. Adaptability ------------ The MQ policy maintains a hit count for each cache block. For a different block to get promoted to the cache it's hit count has to exceed the lowest currently in the cache. This means it can take a long time for the cache to adapt between varying IO patterns. Periodically degrading the hit counts could help with this, but I haven't found a nice general solution. SMQ doesn't maintain hit counts, so a lot of this problem just goes away. In addition it tracks performance of the hotspot queue, which is used to decide which blocks to promote. If the hotspot queue is performing badly then it starts moving entries more quickly between levels. This lets it adapt to new IO patterns very quickly. Performance ----------- In my tests SMQ shows substantially better performance than MQ. Once this matures a bit more I'm sure it'll become the default policy. Signed-off-by: Joe Thornber <ejt@redhat.com> Signed-off-by: Mike Snitzer <snitzer@redhat.com>
2015-05-15 14:33:34 +00:00
__smq_set_clear_dirty(mq, oblock, true);
spin_unlock_irqrestore(&mq->lock, flags);
dm cache: add stochastic-multi-queue (smq) policy The stochastic-multi-queue (smq) policy addresses some of the problems with the current multiqueue (mq) policy. Memory usage ------------ The mq policy uses a lot of memory; 88 bytes per cache block on a 64 bit machine. SMQ uses 28bit indexes to implement it's data structures rather than pointers. It avoids storing an explicit hit count for each block. It has a 'hotspot' queue rather than a pre cache which uses a quarter of the entries (each hotspot block covers a larger area than a single cache block). All these mean smq uses ~25bytes per cache block. Still a lot of memory, but a substantial improvement nontheless. Level balancing --------------- MQ places entries in different levels of the multiqueue structures based on their hit count (~ln(hit count)). This means the bottom levels generally have the most entries, and the top ones have very few. Having unbalanced levels like this reduces the efficacy of the multiqueue. SMQ does not maintain a hit count, instead it swaps hit entries with the least recently used entry from the level above. The over all ordering being a side effect of this stochastic process. With this scheme we can decide how many entries occupy each multiqueue level, resulting in better promotion/demotion decisions. Adaptability ------------ The MQ policy maintains a hit count for each cache block. For a different block to get promoted to the cache it's hit count has to exceed the lowest currently in the cache. This means it can take a long time for the cache to adapt between varying IO patterns. Periodically degrading the hit counts could help with this, but I haven't found a nice general solution. SMQ doesn't maintain hit counts, so a lot of this problem just goes away. In addition it tracks performance of the hotspot queue, which is used to decide which blocks to promote. If the hotspot queue is performing badly then it starts moving entries more quickly between levels. This lets it adapt to new IO patterns very quickly. Performance ----------- In my tests SMQ shows substantially better performance than MQ. Once this matures a bit more I'm sure it'll become the default policy. Signed-off-by: Joe Thornber <ejt@redhat.com> Signed-off-by: Mike Snitzer <snitzer@redhat.com>
2015-05-15 14:33:34 +00:00
}
static void smq_clear_dirty(struct dm_cache_policy *p, dm_oblock_t oblock)
{
struct smq_policy *mq = to_smq_policy(p);
unsigned long flags;
dm cache: add stochastic-multi-queue (smq) policy The stochastic-multi-queue (smq) policy addresses some of the problems with the current multiqueue (mq) policy. Memory usage ------------ The mq policy uses a lot of memory; 88 bytes per cache block on a 64 bit machine. SMQ uses 28bit indexes to implement it's data structures rather than pointers. It avoids storing an explicit hit count for each block. It has a 'hotspot' queue rather than a pre cache which uses a quarter of the entries (each hotspot block covers a larger area than a single cache block). All these mean smq uses ~25bytes per cache block. Still a lot of memory, but a substantial improvement nontheless. Level balancing --------------- MQ places entries in different levels of the multiqueue structures based on their hit count (~ln(hit count)). This means the bottom levels generally have the most entries, and the top ones have very few. Having unbalanced levels like this reduces the efficacy of the multiqueue. SMQ does not maintain a hit count, instead it swaps hit entries with the least recently used entry from the level above. The over all ordering being a side effect of this stochastic process. With this scheme we can decide how many entries occupy each multiqueue level, resulting in better promotion/demotion decisions. Adaptability ------------ The MQ policy maintains a hit count for each cache block. For a different block to get promoted to the cache it's hit count has to exceed the lowest currently in the cache. This means it can take a long time for the cache to adapt between varying IO patterns. Periodically degrading the hit counts could help with this, but I haven't found a nice general solution. SMQ doesn't maintain hit counts, so a lot of this problem just goes away. In addition it tracks performance of the hotspot queue, which is used to decide which blocks to promote. If the hotspot queue is performing badly then it starts moving entries more quickly between levels. This lets it adapt to new IO patterns very quickly. Performance ----------- In my tests SMQ shows substantially better performance than MQ. Once this matures a bit more I'm sure it'll become the default policy. Signed-off-by: Joe Thornber <ejt@redhat.com> Signed-off-by: Mike Snitzer <snitzer@redhat.com>
2015-05-15 14:33:34 +00:00
spin_lock_irqsave(&mq->lock, flags);
dm cache: add stochastic-multi-queue (smq) policy The stochastic-multi-queue (smq) policy addresses some of the problems with the current multiqueue (mq) policy. Memory usage ------------ The mq policy uses a lot of memory; 88 bytes per cache block on a 64 bit machine. SMQ uses 28bit indexes to implement it's data structures rather than pointers. It avoids storing an explicit hit count for each block. It has a 'hotspot' queue rather than a pre cache which uses a quarter of the entries (each hotspot block covers a larger area than a single cache block). All these mean smq uses ~25bytes per cache block. Still a lot of memory, but a substantial improvement nontheless. Level balancing --------------- MQ places entries in different levels of the multiqueue structures based on their hit count (~ln(hit count)). This means the bottom levels generally have the most entries, and the top ones have very few. Having unbalanced levels like this reduces the efficacy of the multiqueue. SMQ does not maintain a hit count, instead it swaps hit entries with the least recently used entry from the level above. The over all ordering being a side effect of this stochastic process. With this scheme we can decide how many entries occupy each multiqueue level, resulting in better promotion/demotion decisions. Adaptability ------------ The MQ policy maintains a hit count for each cache block. For a different block to get promoted to the cache it's hit count has to exceed the lowest currently in the cache. This means it can take a long time for the cache to adapt between varying IO patterns. Periodically degrading the hit counts could help with this, but I haven't found a nice general solution. SMQ doesn't maintain hit counts, so a lot of this problem just goes away. In addition it tracks performance of the hotspot queue, which is used to decide which blocks to promote. If the hotspot queue is performing badly then it starts moving entries more quickly between levels. This lets it adapt to new IO patterns very quickly. Performance ----------- In my tests SMQ shows substantially better performance than MQ. Once this matures a bit more I'm sure it'll become the default policy. Signed-off-by: Joe Thornber <ejt@redhat.com> Signed-off-by: Mike Snitzer <snitzer@redhat.com>
2015-05-15 14:33:34 +00:00
__smq_set_clear_dirty(mq, oblock, false);
spin_unlock_irqrestore(&mq->lock, flags);
dm cache: add stochastic-multi-queue (smq) policy The stochastic-multi-queue (smq) policy addresses some of the problems with the current multiqueue (mq) policy. Memory usage ------------ The mq policy uses a lot of memory; 88 bytes per cache block on a 64 bit machine. SMQ uses 28bit indexes to implement it's data structures rather than pointers. It avoids storing an explicit hit count for each block. It has a 'hotspot' queue rather than a pre cache which uses a quarter of the entries (each hotspot block covers a larger area than a single cache block). All these mean smq uses ~25bytes per cache block. Still a lot of memory, but a substantial improvement nontheless. Level balancing --------------- MQ places entries in different levels of the multiqueue structures based on their hit count (~ln(hit count)). This means the bottom levels generally have the most entries, and the top ones have very few. Having unbalanced levels like this reduces the efficacy of the multiqueue. SMQ does not maintain a hit count, instead it swaps hit entries with the least recently used entry from the level above. The over all ordering being a side effect of this stochastic process. With this scheme we can decide how many entries occupy each multiqueue level, resulting in better promotion/demotion decisions. Adaptability ------------ The MQ policy maintains a hit count for each cache block. For a different block to get promoted to the cache it's hit count has to exceed the lowest currently in the cache. This means it can take a long time for the cache to adapt between varying IO patterns. Periodically degrading the hit counts could help with this, but I haven't found a nice general solution. SMQ doesn't maintain hit counts, so a lot of this problem just goes away. In addition it tracks performance of the hotspot queue, which is used to decide which blocks to promote. If the hotspot queue is performing badly then it starts moving entries more quickly between levels. This lets it adapt to new IO patterns very quickly. Performance ----------- In my tests SMQ shows substantially better performance than MQ. Once this matures a bit more I'm sure it'll become the default policy. Signed-off-by: Joe Thornber <ejt@redhat.com> Signed-off-by: Mike Snitzer <snitzer@redhat.com>
2015-05-15 14:33:34 +00:00
}
static int smq_load_mapping(struct dm_cache_policy *p,
dm_oblock_t oblock, dm_cblock_t cblock,
uint32_t hint, bool hint_valid)
{
struct smq_policy *mq = to_smq_policy(p);
struct entry *e;
e = alloc_particular_entry(&mq->cache_alloc, from_cblock(cblock));
e->oblock = oblock;
e->dirty = false; /* this gets corrected in a minute */
e->level = hint_valid ? min(hint, NR_CACHE_LEVELS - 1) : 1;
push(mq, e);
return 0;
}
static int smq_save_hints(struct smq_policy *mq, struct queue *q,
policy_walk_fn fn, void *context)
{
int r;
unsigned level;
struct entry *e;
for (level = 0; level < q->nr_levels; level++)
for (e = l_head(q->es, q->qs + level); e; e = l_next(q->es, e)) {
if (!e->sentinel) {
r = fn(context, infer_cblock(mq, e),
e->oblock, e->level);
if (r)
return r;
}
}
return 0;
}
static int smq_walk_mappings(struct dm_cache_policy *p, policy_walk_fn fn,
void *context)
{
struct smq_policy *mq = to_smq_policy(p);
int r = 0;
/*
* We don't need to lock here since this method is only called once
* the IO has stopped.
*/
dm cache: add stochastic-multi-queue (smq) policy The stochastic-multi-queue (smq) policy addresses some of the problems with the current multiqueue (mq) policy. Memory usage ------------ The mq policy uses a lot of memory; 88 bytes per cache block on a 64 bit machine. SMQ uses 28bit indexes to implement it's data structures rather than pointers. It avoids storing an explicit hit count for each block. It has a 'hotspot' queue rather than a pre cache which uses a quarter of the entries (each hotspot block covers a larger area than a single cache block). All these mean smq uses ~25bytes per cache block. Still a lot of memory, but a substantial improvement nontheless. Level balancing --------------- MQ places entries in different levels of the multiqueue structures based on their hit count (~ln(hit count)). This means the bottom levels generally have the most entries, and the top ones have very few. Having unbalanced levels like this reduces the efficacy of the multiqueue. SMQ does not maintain a hit count, instead it swaps hit entries with the least recently used entry from the level above. The over all ordering being a side effect of this stochastic process. With this scheme we can decide how many entries occupy each multiqueue level, resulting in better promotion/demotion decisions. Adaptability ------------ The MQ policy maintains a hit count for each cache block. For a different block to get promoted to the cache it's hit count has to exceed the lowest currently in the cache. This means it can take a long time for the cache to adapt between varying IO patterns. Periodically degrading the hit counts could help with this, but I haven't found a nice general solution. SMQ doesn't maintain hit counts, so a lot of this problem just goes away. In addition it tracks performance of the hotspot queue, which is used to decide which blocks to promote. If the hotspot queue is performing badly then it starts moving entries more quickly between levels. This lets it adapt to new IO patterns very quickly. Performance ----------- In my tests SMQ shows substantially better performance than MQ. Once this matures a bit more I'm sure it'll become the default policy. Signed-off-by: Joe Thornber <ejt@redhat.com> Signed-off-by: Mike Snitzer <snitzer@redhat.com>
2015-05-15 14:33:34 +00:00
r = smq_save_hints(mq, &mq->clean, fn, context);
if (!r)
r = smq_save_hints(mq, &mq->dirty, fn, context);
return r;
}
static void __remove_mapping(struct smq_policy *mq, dm_oblock_t oblock)
{
struct entry *e;
e = h_lookup(&mq->table, oblock);
BUG_ON(!e);
del(mq, e);
free_entry(&mq->cache_alloc, e);
}
static void smq_remove_mapping(struct dm_cache_policy *p, dm_oblock_t oblock)
{
struct smq_policy *mq = to_smq_policy(p);
unsigned long flags;
dm cache: add stochastic-multi-queue (smq) policy The stochastic-multi-queue (smq) policy addresses some of the problems with the current multiqueue (mq) policy. Memory usage ------------ The mq policy uses a lot of memory; 88 bytes per cache block on a 64 bit machine. SMQ uses 28bit indexes to implement it's data structures rather than pointers. It avoids storing an explicit hit count for each block. It has a 'hotspot' queue rather than a pre cache which uses a quarter of the entries (each hotspot block covers a larger area than a single cache block). All these mean smq uses ~25bytes per cache block. Still a lot of memory, but a substantial improvement nontheless. Level balancing --------------- MQ places entries in different levels of the multiqueue structures based on their hit count (~ln(hit count)). This means the bottom levels generally have the most entries, and the top ones have very few. Having unbalanced levels like this reduces the efficacy of the multiqueue. SMQ does not maintain a hit count, instead it swaps hit entries with the least recently used entry from the level above. The over all ordering being a side effect of this stochastic process. With this scheme we can decide how many entries occupy each multiqueue level, resulting in better promotion/demotion decisions. Adaptability ------------ The MQ policy maintains a hit count for each cache block. For a different block to get promoted to the cache it's hit count has to exceed the lowest currently in the cache. This means it can take a long time for the cache to adapt between varying IO patterns. Periodically degrading the hit counts could help with this, but I haven't found a nice general solution. SMQ doesn't maintain hit counts, so a lot of this problem just goes away. In addition it tracks performance of the hotspot queue, which is used to decide which blocks to promote. If the hotspot queue is performing badly then it starts moving entries more quickly between levels. This lets it adapt to new IO patterns very quickly. Performance ----------- In my tests SMQ shows substantially better performance than MQ. Once this matures a bit more I'm sure it'll become the default policy. Signed-off-by: Joe Thornber <ejt@redhat.com> Signed-off-by: Mike Snitzer <snitzer@redhat.com>
2015-05-15 14:33:34 +00:00
spin_lock_irqsave(&mq->lock, flags);
dm cache: add stochastic-multi-queue (smq) policy The stochastic-multi-queue (smq) policy addresses some of the problems with the current multiqueue (mq) policy. Memory usage ------------ The mq policy uses a lot of memory; 88 bytes per cache block on a 64 bit machine. SMQ uses 28bit indexes to implement it's data structures rather than pointers. It avoids storing an explicit hit count for each block. It has a 'hotspot' queue rather than a pre cache which uses a quarter of the entries (each hotspot block covers a larger area than a single cache block). All these mean smq uses ~25bytes per cache block. Still a lot of memory, but a substantial improvement nontheless. Level balancing --------------- MQ places entries in different levels of the multiqueue structures based on their hit count (~ln(hit count)). This means the bottom levels generally have the most entries, and the top ones have very few. Having unbalanced levels like this reduces the efficacy of the multiqueue. SMQ does not maintain a hit count, instead it swaps hit entries with the least recently used entry from the level above. The over all ordering being a side effect of this stochastic process. With this scheme we can decide how many entries occupy each multiqueue level, resulting in better promotion/demotion decisions. Adaptability ------------ The MQ policy maintains a hit count for each cache block. For a different block to get promoted to the cache it's hit count has to exceed the lowest currently in the cache. This means it can take a long time for the cache to adapt between varying IO patterns. Periodically degrading the hit counts could help with this, but I haven't found a nice general solution. SMQ doesn't maintain hit counts, so a lot of this problem just goes away. In addition it tracks performance of the hotspot queue, which is used to decide which blocks to promote. If the hotspot queue is performing badly then it starts moving entries more quickly between levels. This lets it adapt to new IO patterns very quickly. Performance ----------- In my tests SMQ shows substantially better performance than MQ. Once this matures a bit more I'm sure it'll become the default policy. Signed-off-by: Joe Thornber <ejt@redhat.com> Signed-off-by: Mike Snitzer <snitzer@redhat.com>
2015-05-15 14:33:34 +00:00
__remove_mapping(mq, oblock);
spin_unlock_irqrestore(&mq->lock, flags);
dm cache: add stochastic-multi-queue (smq) policy The stochastic-multi-queue (smq) policy addresses some of the problems with the current multiqueue (mq) policy. Memory usage ------------ The mq policy uses a lot of memory; 88 bytes per cache block on a 64 bit machine. SMQ uses 28bit indexes to implement it's data structures rather than pointers. It avoids storing an explicit hit count for each block. It has a 'hotspot' queue rather than a pre cache which uses a quarter of the entries (each hotspot block covers a larger area than a single cache block). All these mean smq uses ~25bytes per cache block. Still a lot of memory, but a substantial improvement nontheless. Level balancing --------------- MQ places entries in different levels of the multiqueue structures based on their hit count (~ln(hit count)). This means the bottom levels generally have the most entries, and the top ones have very few. Having unbalanced levels like this reduces the efficacy of the multiqueue. SMQ does not maintain a hit count, instead it swaps hit entries with the least recently used entry from the level above. The over all ordering being a side effect of this stochastic process. With this scheme we can decide how many entries occupy each multiqueue level, resulting in better promotion/demotion decisions. Adaptability ------------ The MQ policy maintains a hit count for each cache block. For a different block to get promoted to the cache it's hit count has to exceed the lowest currently in the cache. This means it can take a long time for the cache to adapt between varying IO patterns. Periodically degrading the hit counts could help with this, but I haven't found a nice general solution. SMQ doesn't maintain hit counts, so a lot of this problem just goes away. In addition it tracks performance of the hotspot queue, which is used to decide which blocks to promote. If the hotspot queue is performing badly then it starts moving entries more quickly between levels. This lets it adapt to new IO patterns very quickly. Performance ----------- In my tests SMQ shows substantially better performance than MQ. Once this matures a bit more I'm sure it'll become the default policy. Signed-off-by: Joe Thornber <ejt@redhat.com> Signed-off-by: Mike Snitzer <snitzer@redhat.com>
2015-05-15 14:33:34 +00:00
}
static int __remove_cblock(struct smq_policy *mq, dm_cblock_t cblock)
{
struct entry *e = get_entry(&mq->cache_alloc, from_cblock(cblock));
if (!e || !e->allocated)
return -ENODATA;
del(mq, e);
free_entry(&mq->cache_alloc, e);
return 0;
}
static int smq_remove_cblock(struct dm_cache_policy *p, dm_cblock_t cblock)
{
int r;
unsigned long flags;
dm cache: add stochastic-multi-queue (smq) policy The stochastic-multi-queue (smq) policy addresses some of the problems with the current multiqueue (mq) policy. Memory usage ------------ The mq policy uses a lot of memory; 88 bytes per cache block on a 64 bit machine. SMQ uses 28bit indexes to implement it's data structures rather than pointers. It avoids storing an explicit hit count for each block. It has a 'hotspot' queue rather than a pre cache which uses a quarter of the entries (each hotspot block covers a larger area than a single cache block). All these mean smq uses ~25bytes per cache block. Still a lot of memory, but a substantial improvement nontheless. Level balancing --------------- MQ places entries in different levels of the multiqueue structures based on their hit count (~ln(hit count)). This means the bottom levels generally have the most entries, and the top ones have very few. Having unbalanced levels like this reduces the efficacy of the multiqueue. SMQ does not maintain a hit count, instead it swaps hit entries with the least recently used entry from the level above. The over all ordering being a side effect of this stochastic process. With this scheme we can decide how many entries occupy each multiqueue level, resulting in better promotion/demotion decisions. Adaptability ------------ The MQ policy maintains a hit count for each cache block. For a different block to get promoted to the cache it's hit count has to exceed the lowest currently in the cache. This means it can take a long time for the cache to adapt between varying IO patterns. Periodically degrading the hit counts could help with this, but I haven't found a nice general solution. SMQ doesn't maintain hit counts, so a lot of this problem just goes away. In addition it tracks performance of the hotspot queue, which is used to decide which blocks to promote. If the hotspot queue is performing badly then it starts moving entries more quickly between levels. This lets it adapt to new IO patterns very quickly. Performance ----------- In my tests SMQ shows substantially better performance than MQ. Once this matures a bit more I'm sure it'll become the default policy. Signed-off-by: Joe Thornber <ejt@redhat.com> Signed-off-by: Mike Snitzer <snitzer@redhat.com>
2015-05-15 14:33:34 +00:00
struct smq_policy *mq = to_smq_policy(p);
spin_lock_irqsave(&mq->lock, flags);
dm cache: add stochastic-multi-queue (smq) policy The stochastic-multi-queue (smq) policy addresses some of the problems with the current multiqueue (mq) policy. Memory usage ------------ The mq policy uses a lot of memory; 88 bytes per cache block on a 64 bit machine. SMQ uses 28bit indexes to implement it's data structures rather than pointers. It avoids storing an explicit hit count for each block. It has a 'hotspot' queue rather than a pre cache which uses a quarter of the entries (each hotspot block covers a larger area than a single cache block). All these mean smq uses ~25bytes per cache block. Still a lot of memory, but a substantial improvement nontheless. Level balancing --------------- MQ places entries in different levels of the multiqueue structures based on their hit count (~ln(hit count)). This means the bottom levels generally have the most entries, and the top ones have very few. Having unbalanced levels like this reduces the efficacy of the multiqueue. SMQ does not maintain a hit count, instead it swaps hit entries with the least recently used entry from the level above. The over all ordering being a side effect of this stochastic process. With this scheme we can decide how many entries occupy each multiqueue level, resulting in better promotion/demotion decisions. Adaptability ------------ The MQ policy maintains a hit count for each cache block. For a different block to get promoted to the cache it's hit count has to exceed the lowest currently in the cache. This means it can take a long time for the cache to adapt between varying IO patterns. Periodically degrading the hit counts could help with this, but I haven't found a nice general solution. SMQ doesn't maintain hit counts, so a lot of this problem just goes away. In addition it tracks performance of the hotspot queue, which is used to decide which blocks to promote. If the hotspot queue is performing badly then it starts moving entries more quickly between levels. This lets it adapt to new IO patterns very quickly. Performance ----------- In my tests SMQ shows substantially better performance than MQ. Once this matures a bit more I'm sure it'll become the default policy. Signed-off-by: Joe Thornber <ejt@redhat.com> Signed-off-by: Mike Snitzer <snitzer@redhat.com>
2015-05-15 14:33:34 +00:00
r = __remove_cblock(mq, cblock);
spin_unlock_irqrestore(&mq->lock, flags);
dm cache: add stochastic-multi-queue (smq) policy The stochastic-multi-queue (smq) policy addresses some of the problems with the current multiqueue (mq) policy. Memory usage ------------ The mq policy uses a lot of memory; 88 bytes per cache block on a 64 bit machine. SMQ uses 28bit indexes to implement it's data structures rather than pointers. It avoids storing an explicit hit count for each block. It has a 'hotspot' queue rather than a pre cache which uses a quarter of the entries (each hotspot block covers a larger area than a single cache block). All these mean smq uses ~25bytes per cache block. Still a lot of memory, but a substantial improvement nontheless. Level balancing --------------- MQ places entries in different levels of the multiqueue structures based on their hit count (~ln(hit count)). This means the bottom levels generally have the most entries, and the top ones have very few. Having unbalanced levels like this reduces the efficacy of the multiqueue. SMQ does not maintain a hit count, instead it swaps hit entries with the least recently used entry from the level above. The over all ordering being a side effect of this stochastic process. With this scheme we can decide how many entries occupy each multiqueue level, resulting in better promotion/demotion decisions. Adaptability ------------ The MQ policy maintains a hit count for each cache block. For a different block to get promoted to the cache it's hit count has to exceed the lowest currently in the cache. This means it can take a long time for the cache to adapt between varying IO patterns. Periodically degrading the hit counts could help with this, but I haven't found a nice general solution. SMQ doesn't maintain hit counts, so a lot of this problem just goes away. In addition it tracks performance of the hotspot queue, which is used to decide which blocks to promote. If the hotspot queue is performing badly then it starts moving entries more quickly between levels. This lets it adapt to new IO patterns very quickly. Performance ----------- In my tests SMQ shows substantially better performance than MQ. Once this matures a bit more I'm sure it'll become the default policy. Signed-off-by: Joe Thornber <ejt@redhat.com> Signed-off-by: Mike Snitzer <snitzer@redhat.com>
2015-05-15 14:33:34 +00:00
return r;
}
#define CLEAN_TARGET_CRITICAL 5u /* percent */
static bool clean_target_met(struct smq_policy *mq, bool critical)
{
if (critical) {
/*
* Cache entries may not be populated. So we're cannot rely on the
* size of the clean queue.
*/
unsigned nr_clean = from_cblock(mq->cache_size) - q_size(&mq->dirty);
unsigned target = from_cblock(mq->cache_size) * CLEAN_TARGET_CRITICAL / 100u;
return nr_clean >= target;
} else
return !q_size(&mq->dirty);
}
static int __smq_writeback_work(struct smq_policy *mq, dm_oblock_t *oblock,
dm_cblock_t *cblock, bool critical_only)
{
struct entry *e = NULL;
bool target_met = clean_target_met(mq, critical_only);
if (critical_only)
/*
* Always try and keep the bottom level clean.
*/
e = pop_old(mq, &mq->dirty, target_met ? 1u : mq->dirty.nr_levels);
else
e = pop_old(mq, &mq->dirty, mq->dirty.nr_levels);
if (!e)
return -ENODATA;
*oblock = e->oblock;
*cblock = infer_cblock(mq, e);
e->dirty = false;
push_new(mq, e);
return 0;
}
static int smq_writeback_work(struct dm_cache_policy *p, dm_oblock_t *oblock,
dm_cblock_t *cblock, bool critical_only)
{
int r;
unsigned long flags;
dm cache: add stochastic-multi-queue (smq) policy The stochastic-multi-queue (smq) policy addresses some of the problems with the current multiqueue (mq) policy. Memory usage ------------ The mq policy uses a lot of memory; 88 bytes per cache block on a 64 bit machine. SMQ uses 28bit indexes to implement it's data structures rather than pointers. It avoids storing an explicit hit count for each block. It has a 'hotspot' queue rather than a pre cache which uses a quarter of the entries (each hotspot block covers a larger area than a single cache block). All these mean smq uses ~25bytes per cache block. Still a lot of memory, but a substantial improvement nontheless. Level balancing --------------- MQ places entries in different levels of the multiqueue structures based on their hit count (~ln(hit count)). This means the bottom levels generally have the most entries, and the top ones have very few. Having unbalanced levels like this reduces the efficacy of the multiqueue. SMQ does not maintain a hit count, instead it swaps hit entries with the least recently used entry from the level above. The over all ordering being a side effect of this stochastic process. With this scheme we can decide how many entries occupy each multiqueue level, resulting in better promotion/demotion decisions. Adaptability ------------ The MQ policy maintains a hit count for each cache block. For a different block to get promoted to the cache it's hit count has to exceed the lowest currently in the cache. This means it can take a long time for the cache to adapt between varying IO patterns. Periodically degrading the hit counts could help with this, but I haven't found a nice general solution. SMQ doesn't maintain hit counts, so a lot of this problem just goes away. In addition it tracks performance of the hotspot queue, which is used to decide which blocks to promote. If the hotspot queue is performing badly then it starts moving entries more quickly between levels. This lets it adapt to new IO patterns very quickly. Performance ----------- In my tests SMQ shows substantially better performance than MQ. Once this matures a bit more I'm sure it'll become the default policy. Signed-off-by: Joe Thornber <ejt@redhat.com> Signed-off-by: Mike Snitzer <snitzer@redhat.com>
2015-05-15 14:33:34 +00:00
struct smq_policy *mq = to_smq_policy(p);
spin_lock_irqsave(&mq->lock, flags);
dm cache: add stochastic-multi-queue (smq) policy The stochastic-multi-queue (smq) policy addresses some of the problems with the current multiqueue (mq) policy. Memory usage ------------ The mq policy uses a lot of memory; 88 bytes per cache block on a 64 bit machine. SMQ uses 28bit indexes to implement it's data structures rather than pointers. It avoids storing an explicit hit count for each block. It has a 'hotspot' queue rather than a pre cache which uses a quarter of the entries (each hotspot block covers a larger area than a single cache block). All these mean smq uses ~25bytes per cache block. Still a lot of memory, but a substantial improvement nontheless. Level balancing --------------- MQ places entries in different levels of the multiqueue structures based on their hit count (~ln(hit count)). This means the bottom levels generally have the most entries, and the top ones have very few. Having unbalanced levels like this reduces the efficacy of the multiqueue. SMQ does not maintain a hit count, instead it swaps hit entries with the least recently used entry from the level above. The over all ordering being a side effect of this stochastic process. With this scheme we can decide how many entries occupy each multiqueue level, resulting in better promotion/demotion decisions. Adaptability ------------ The MQ policy maintains a hit count for each cache block. For a different block to get promoted to the cache it's hit count has to exceed the lowest currently in the cache. This means it can take a long time for the cache to adapt between varying IO patterns. Periodically degrading the hit counts could help with this, but I haven't found a nice general solution. SMQ doesn't maintain hit counts, so a lot of this problem just goes away. In addition it tracks performance of the hotspot queue, which is used to decide which blocks to promote. If the hotspot queue is performing badly then it starts moving entries more quickly between levels. This lets it adapt to new IO patterns very quickly. Performance ----------- In my tests SMQ shows substantially better performance than MQ. Once this matures a bit more I'm sure it'll become the default policy. Signed-off-by: Joe Thornber <ejt@redhat.com> Signed-off-by: Mike Snitzer <snitzer@redhat.com>
2015-05-15 14:33:34 +00:00
r = __smq_writeback_work(mq, oblock, cblock, critical_only);
spin_unlock_irqrestore(&mq->lock, flags);
dm cache: add stochastic-multi-queue (smq) policy The stochastic-multi-queue (smq) policy addresses some of the problems with the current multiqueue (mq) policy. Memory usage ------------ The mq policy uses a lot of memory; 88 bytes per cache block on a 64 bit machine. SMQ uses 28bit indexes to implement it's data structures rather than pointers. It avoids storing an explicit hit count for each block. It has a 'hotspot' queue rather than a pre cache which uses a quarter of the entries (each hotspot block covers a larger area than a single cache block). All these mean smq uses ~25bytes per cache block. Still a lot of memory, but a substantial improvement nontheless. Level balancing --------------- MQ places entries in different levels of the multiqueue structures based on their hit count (~ln(hit count)). This means the bottom levels generally have the most entries, and the top ones have very few. Having unbalanced levels like this reduces the efficacy of the multiqueue. SMQ does not maintain a hit count, instead it swaps hit entries with the least recently used entry from the level above. The over all ordering being a side effect of this stochastic process. With this scheme we can decide how many entries occupy each multiqueue level, resulting in better promotion/demotion decisions. Adaptability ------------ The MQ policy maintains a hit count for each cache block. For a different block to get promoted to the cache it's hit count has to exceed the lowest currently in the cache. This means it can take a long time for the cache to adapt between varying IO patterns. Periodically degrading the hit counts could help with this, but I haven't found a nice general solution. SMQ doesn't maintain hit counts, so a lot of this problem just goes away. In addition it tracks performance of the hotspot queue, which is used to decide which blocks to promote. If the hotspot queue is performing badly then it starts moving entries more quickly between levels. This lets it adapt to new IO patterns very quickly. Performance ----------- In my tests SMQ shows substantially better performance than MQ. Once this matures a bit more I'm sure it'll become the default policy. Signed-off-by: Joe Thornber <ejt@redhat.com> Signed-off-by: Mike Snitzer <snitzer@redhat.com>
2015-05-15 14:33:34 +00:00
return r;
}
static void __force_mapping(struct smq_policy *mq,
dm_oblock_t current_oblock, dm_oblock_t new_oblock)
{
struct entry *e = h_lookup(&mq->table, current_oblock);
if (e) {
del(mq, e);
e->oblock = new_oblock;
e->dirty = true;
push(mq, e);
}
}
static void smq_force_mapping(struct dm_cache_policy *p,
dm_oblock_t current_oblock, dm_oblock_t new_oblock)
{
unsigned long flags;
dm cache: add stochastic-multi-queue (smq) policy The stochastic-multi-queue (smq) policy addresses some of the problems with the current multiqueue (mq) policy. Memory usage ------------ The mq policy uses a lot of memory; 88 bytes per cache block on a 64 bit machine. SMQ uses 28bit indexes to implement it's data structures rather than pointers. It avoids storing an explicit hit count for each block. It has a 'hotspot' queue rather than a pre cache which uses a quarter of the entries (each hotspot block covers a larger area than a single cache block). All these mean smq uses ~25bytes per cache block. Still a lot of memory, but a substantial improvement nontheless. Level balancing --------------- MQ places entries in different levels of the multiqueue structures based on their hit count (~ln(hit count)). This means the bottom levels generally have the most entries, and the top ones have very few. Having unbalanced levels like this reduces the efficacy of the multiqueue. SMQ does not maintain a hit count, instead it swaps hit entries with the least recently used entry from the level above. The over all ordering being a side effect of this stochastic process. With this scheme we can decide how many entries occupy each multiqueue level, resulting in better promotion/demotion decisions. Adaptability ------------ The MQ policy maintains a hit count for each cache block. For a different block to get promoted to the cache it's hit count has to exceed the lowest currently in the cache. This means it can take a long time for the cache to adapt between varying IO patterns. Periodically degrading the hit counts could help with this, but I haven't found a nice general solution. SMQ doesn't maintain hit counts, so a lot of this problem just goes away. In addition it tracks performance of the hotspot queue, which is used to decide which blocks to promote. If the hotspot queue is performing badly then it starts moving entries more quickly between levels. This lets it adapt to new IO patterns very quickly. Performance ----------- In my tests SMQ shows substantially better performance than MQ. Once this matures a bit more I'm sure it'll become the default policy. Signed-off-by: Joe Thornber <ejt@redhat.com> Signed-off-by: Mike Snitzer <snitzer@redhat.com>
2015-05-15 14:33:34 +00:00
struct smq_policy *mq = to_smq_policy(p);
spin_lock_irqsave(&mq->lock, flags);
dm cache: add stochastic-multi-queue (smq) policy The stochastic-multi-queue (smq) policy addresses some of the problems with the current multiqueue (mq) policy. Memory usage ------------ The mq policy uses a lot of memory; 88 bytes per cache block on a 64 bit machine. SMQ uses 28bit indexes to implement it's data structures rather than pointers. It avoids storing an explicit hit count for each block. It has a 'hotspot' queue rather than a pre cache which uses a quarter of the entries (each hotspot block covers a larger area than a single cache block). All these mean smq uses ~25bytes per cache block. Still a lot of memory, but a substantial improvement nontheless. Level balancing --------------- MQ places entries in different levels of the multiqueue structures based on their hit count (~ln(hit count)). This means the bottom levels generally have the most entries, and the top ones have very few. Having unbalanced levels like this reduces the efficacy of the multiqueue. SMQ does not maintain a hit count, instead it swaps hit entries with the least recently used entry from the level above. The over all ordering being a side effect of this stochastic process. With this scheme we can decide how many entries occupy each multiqueue level, resulting in better promotion/demotion decisions. Adaptability ------------ The MQ policy maintains a hit count for each cache block. For a different block to get promoted to the cache it's hit count has to exceed the lowest currently in the cache. This means it can take a long time for the cache to adapt between varying IO patterns. Periodically degrading the hit counts could help with this, but I haven't found a nice general solution. SMQ doesn't maintain hit counts, so a lot of this problem just goes away. In addition it tracks performance of the hotspot queue, which is used to decide which blocks to promote. If the hotspot queue is performing badly then it starts moving entries more quickly between levels. This lets it adapt to new IO patterns very quickly. Performance ----------- In my tests SMQ shows substantially better performance than MQ. Once this matures a bit more I'm sure it'll become the default policy. Signed-off-by: Joe Thornber <ejt@redhat.com> Signed-off-by: Mike Snitzer <snitzer@redhat.com>
2015-05-15 14:33:34 +00:00
__force_mapping(mq, current_oblock, new_oblock);
spin_unlock_irqrestore(&mq->lock, flags);
dm cache: add stochastic-multi-queue (smq) policy The stochastic-multi-queue (smq) policy addresses some of the problems with the current multiqueue (mq) policy. Memory usage ------------ The mq policy uses a lot of memory; 88 bytes per cache block on a 64 bit machine. SMQ uses 28bit indexes to implement it's data structures rather than pointers. It avoids storing an explicit hit count for each block. It has a 'hotspot' queue rather than a pre cache which uses a quarter of the entries (each hotspot block covers a larger area than a single cache block). All these mean smq uses ~25bytes per cache block. Still a lot of memory, but a substantial improvement nontheless. Level balancing --------------- MQ places entries in different levels of the multiqueue structures based on their hit count (~ln(hit count)). This means the bottom levels generally have the most entries, and the top ones have very few. Having unbalanced levels like this reduces the efficacy of the multiqueue. SMQ does not maintain a hit count, instead it swaps hit entries with the least recently used entry from the level above. The over all ordering being a side effect of this stochastic process. With this scheme we can decide how many entries occupy each multiqueue level, resulting in better promotion/demotion decisions. Adaptability ------------ The MQ policy maintains a hit count for each cache block. For a different block to get promoted to the cache it's hit count has to exceed the lowest currently in the cache. This means it can take a long time for the cache to adapt between varying IO patterns. Periodically degrading the hit counts could help with this, but I haven't found a nice general solution. SMQ doesn't maintain hit counts, so a lot of this problem just goes away. In addition it tracks performance of the hotspot queue, which is used to decide which blocks to promote. If the hotspot queue is performing badly then it starts moving entries more quickly between levels. This lets it adapt to new IO patterns very quickly. Performance ----------- In my tests SMQ shows substantially better performance than MQ. Once this matures a bit more I'm sure it'll become the default policy. Signed-off-by: Joe Thornber <ejt@redhat.com> Signed-off-by: Mike Snitzer <snitzer@redhat.com>
2015-05-15 14:33:34 +00:00
}
static dm_cblock_t smq_residency(struct dm_cache_policy *p)
{
dm_cblock_t r;
unsigned long flags;
dm cache: add stochastic-multi-queue (smq) policy The stochastic-multi-queue (smq) policy addresses some of the problems with the current multiqueue (mq) policy. Memory usage ------------ The mq policy uses a lot of memory; 88 bytes per cache block on a 64 bit machine. SMQ uses 28bit indexes to implement it's data structures rather than pointers. It avoids storing an explicit hit count for each block. It has a 'hotspot' queue rather than a pre cache which uses a quarter of the entries (each hotspot block covers a larger area than a single cache block). All these mean smq uses ~25bytes per cache block. Still a lot of memory, but a substantial improvement nontheless. Level balancing --------------- MQ places entries in different levels of the multiqueue structures based on their hit count (~ln(hit count)). This means the bottom levels generally have the most entries, and the top ones have very few. Having unbalanced levels like this reduces the efficacy of the multiqueue. SMQ does not maintain a hit count, instead it swaps hit entries with the least recently used entry from the level above. The over all ordering being a side effect of this stochastic process. With this scheme we can decide how many entries occupy each multiqueue level, resulting in better promotion/demotion decisions. Adaptability ------------ The MQ policy maintains a hit count for each cache block. For a different block to get promoted to the cache it's hit count has to exceed the lowest currently in the cache. This means it can take a long time for the cache to adapt between varying IO patterns. Periodically degrading the hit counts could help with this, but I haven't found a nice general solution. SMQ doesn't maintain hit counts, so a lot of this problem just goes away. In addition it tracks performance of the hotspot queue, which is used to decide which blocks to promote. If the hotspot queue is performing badly then it starts moving entries more quickly between levels. This lets it adapt to new IO patterns very quickly. Performance ----------- In my tests SMQ shows substantially better performance than MQ. Once this matures a bit more I'm sure it'll become the default policy. Signed-off-by: Joe Thornber <ejt@redhat.com> Signed-off-by: Mike Snitzer <snitzer@redhat.com>
2015-05-15 14:33:34 +00:00
struct smq_policy *mq = to_smq_policy(p);
spin_lock_irqsave(&mq->lock, flags);
dm cache: add stochastic-multi-queue (smq) policy The stochastic-multi-queue (smq) policy addresses some of the problems with the current multiqueue (mq) policy. Memory usage ------------ The mq policy uses a lot of memory; 88 bytes per cache block on a 64 bit machine. SMQ uses 28bit indexes to implement it's data structures rather than pointers. It avoids storing an explicit hit count for each block. It has a 'hotspot' queue rather than a pre cache which uses a quarter of the entries (each hotspot block covers a larger area than a single cache block). All these mean smq uses ~25bytes per cache block. Still a lot of memory, but a substantial improvement nontheless. Level balancing --------------- MQ places entries in different levels of the multiqueue structures based on their hit count (~ln(hit count)). This means the bottom levels generally have the most entries, and the top ones have very few. Having unbalanced levels like this reduces the efficacy of the multiqueue. SMQ does not maintain a hit count, instead it swaps hit entries with the least recently used entry from the level above. The over all ordering being a side effect of this stochastic process. With this scheme we can decide how many entries occupy each multiqueue level, resulting in better promotion/demotion decisions. Adaptability ------------ The MQ policy maintains a hit count for each cache block. For a different block to get promoted to the cache it's hit count has to exceed the lowest currently in the cache. This means it can take a long time for the cache to adapt between varying IO patterns. Periodically degrading the hit counts could help with this, but I haven't found a nice general solution. SMQ doesn't maintain hit counts, so a lot of this problem just goes away. In addition it tracks performance of the hotspot queue, which is used to decide which blocks to promote. If the hotspot queue is performing badly then it starts moving entries more quickly between levels. This lets it adapt to new IO patterns very quickly. Performance ----------- In my tests SMQ shows substantially better performance than MQ. Once this matures a bit more I'm sure it'll become the default policy. Signed-off-by: Joe Thornber <ejt@redhat.com> Signed-off-by: Mike Snitzer <snitzer@redhat.com>
2015-05-15 14:33:34 +00:00
r = to_cblock(mq->cache_alloc.nr_allocated);
spin_unlock_irqrestore(&mq->lock, flags);
dm cache: add stochastic-multi-queue (smq) policy The stochastic-multi-queue (smq) policy addresses some of the problems with the current multiqueue (mq) policy. Memory usage ------------ The mq policy uses a lot of memory; 88 bytes per cache block on a 64 bit machine. SMQ uses 28bit indexes to implement it's data structures rather than pointers. It avoids storing an explicit hit count for each block. It has a 'hotspot' queue rather than a pre cache which uses a quarter of the entries (each hotspot block covers a larger area than a single cache block). All these mean smq uses ~25bytes per cache block. Still a lot of memory, but a substantial improvement nontheless. Level balancing --------------- MQ places entries in different levels of the multiqueue structures based on their hit count (~ln(hit count)). This means the bottom levels generally have the most entries, and the top ones have very few. Having unbalanced levels like this reduces the efficacy of the multiqueue. SMQ does not maintain a hit count, instead it swaps hit entries with the least recently used entry from the level above. The over all ordering being a side effect of this stochastic process. With this scheme we can decide how many entries occupy each multiqueue level, resulting in better promotion/demotion decisions. Adaptability ------------ The MQ policy maintains a hit count for each cache block. For a different block to get promoted to the cache it's hit count has to exceed the lowest currently in the cache. This means it can take a long time for the cache to adapt between varying IO patterns. Periodically degrading the hit counts could help with this, but I haven't found a nice general solution. SMQ doesn't maintain hit counts, so a lot of this problem just goes away. In addition it tracks performance of the hotspot queue, which is used to decide which blocks to promote. If the hotspot queue is performing badly then it starts moving entries more quickly between levels. This lets it adapt to new IO patterns very quickly. Performance ----------- In my tests SMQ shows substantially better performance than MQ. Once this matures a bit more I'm sure it'll become the default policy. Signed-off-by: Joe Thornber <ejt@redhat.com> Signed-off-by: Mike Snitzer <snitzer@redhat.com>
2015-05-15 14:33:34 +00:00
return r;
}
static void smq_tick(struct dm_cache_policy *p, bool can_block)
dm cache: add stochastic-multi-queue (smq) policy The stochastic-multi-queue (smq) policy addresses some of the problems with the current multiqueue (mq) policy. Memory usage ------------ The mq policy uses a lot of memory; 88 bytes per cache block on a 64 bit machine. SMQ uses 28bit indexes to implement it's data structures rather than pointers. It avoids storing an explicit hit count for each block. It has a 'hotspot' queue rather than a pre cache which uses a quarter of the entries (each hotspot block covers a larger area than a single cache block). All these mean smq uses ~25bytes per cache block. Still a lot of memory, but a substantial improvement nontheless. Level balancing --------------- MQ places entries in different levels of the multiqueue structures based on their hit count (~ln(hit count)). This means the bottom levels generally have the most entries, and the top ones have very few. Having unbalanced levels like this reduces the efficacy of the multiqueue. SMQ does not maintain a hit count, instead it swaps hit entries with the least recently used entry from the level above. The over all ordering being a side effect of this stochastic process. With this scheme we can decide how many entries occupy each multiqueue level, resulting in better promotion/demotion decisions. Adaptability ------------ The MQ policy maintains a hit count for each cache block. For a different block to get promoted to the cache it's hit count has to exceed the lowest currently in the cache. This means it can take a long time for the cache to adapt between varying IO patterns. Periodically degrading the hit counts could help with this, but I haven't found a nice general solution. SMQ doesn't maintain hit counts, so a lot of this problem just goes away. In addition it tracks performance of the hotspot queue, which is used to decide which blocks to promote. If the hotspot queue is performing badly then it starts moving entries more quickly between levels. This lets it adapt to new IO patterns very quickly. Performance ----------- In my tests SMQ shows substantially better performance than MQ. Once this matures a bit more I'm sure it'll become the default policy. Signed-off-by: Joe Thornber <ejt@redhat.com> Signed-off-by: Mike Snitzer <snitzer@redhat.com>
2015-05-15 14:33:34 +00:00
{
struct smq_policy *mq = to_smq_policy(p);
unsigned long flags;
spin_lock_irqsave(&mq->lock, flags);
mq->tick++;
update_sentinels(mq);
end_hotspot_period(mq);
end_cache_period(mq);
spin_unlock_irqrestore(&mq->lock, flags);
dm cache: add stochastic-multi-queue (smq) policy The stochastic-multi-queue (smq) policy addresses some of the problems with the current multiqueue (mq) policy. Memory usage ------------ The mq policy uses a lot of memory; 88 bytes per cache block on a 64 bit machine. SMQ uses 28bit indexes to implement it's data structures rather than pointers. It avoids storing an explicit hit count for each block. It has a 'hotspot' queue rather than a pre cache which uses a quarter of the entries (each hotspot block covers a larger area than a single cache block). All these mean smq uses ~25bytes per cache block. Still a lot of memory, but a substantial improvement nontheless. Level balancing --------------- MQ places entries in different levels of the multiqueue structures based on their hit count (~ln(hit count)). This means the bottom levels generally have the most entries, and the top ones have very few. Having unbalanced levels like this reduces the efficacy of the multiqueue. SMQ does not maintain a hit count, instead it swaps hit entries with the least recently used entry from the level above. The over all ordering being a side effect of this stochastic process. With this scheme we can decide how many entries occupy each multiqueue level, resulting in better promotion/demotion decisions. Adaptability ------------ The MQ policy maintains a hit count for each cache block. For a different block to get promoted to the cache it's hit count has to exceed the lowest currently in the cache. This means it can take a long time for the cache to adapt between varying IO patterns. Periodically degrading the hit counts could help with this, but I haven't found a nice general solution. SMQ doesn't maintain hit counts, so a lot of this problem just goes away. In addition it tracks performance of the hotspot queue, which is used to decide which blocks to promote. If the hotspot queue is performing badly then it starts moving entries more quickly between levels. This lets it adapt to new IO patterns very quickly. Performance ----------- In my tests SMQ shows substantially better performance than MQ. Once this matures a bit more I'm sure it'll become the default policy. Signed-off-by: Joe Thornber <ejt@redhat.com> Signed-off-by: Mike Snitzer <snitzer@redhat.com>
2015-05-15 14:33:34 +00:00
}
/* Init the policy plugin interface function pointers. */
static void init_policy_functions(struct smq_policy *mq)
{
mq->policy.destroy = smq_destroy;
mq->policy.map = smq_map;
mq->policy.lookup = smq_lookup;
mq->policy.set_dirty = smq_set_dirty;
mq->policy.clear_dirty = smq_clear_dirty;
mq->policy.load_mapping = smq_load_mapping;
mq->policy.walk_mappings = smq_walk_mappings;
mq->policy.remove_mapping = smq_remove_mapping;
mq->policy.remove_cblock = smq_remove_cblock;
mq->policy.writeback_work = smq_writeback_work;
mq->policy.force_mapping = smq_force_mapping;
mq->policy.residency = smq_residency;
mq->policy.tick = smq_tick;
}
static bool too_many_hotspot_blocks(sector_t origin_size,
sector_t hotspot_block_size,
unsigned nr_hotspot_blocks)
{
return (hotspot_block_size * nr_hotspot_blocks) > origin_size;
}
static void calc_hotspot_params(sector_t origin_size,
sector_t cache_block_size,
unsigned nr_cache_blocks,
sector_t *hotspot_block_size,
unsigned *nr_hotspot_blocks)
{
*hotspot_block_size = cache_block_size * 16u;
*nr_hotspot_blocks = max(nr_cache_blocks / 4u, 1024u);
while ((*hotspot_block_size > cache_block_size) &&
too_many_hotspot_blocks(origin_size, *hotspot_block_size, *nr_hotspot_blocks))
*hotspot_block_size /= 2u;
}
static struct dm_cache_policy *smq_create(dm_cblock_t cache_size,
sector_t origin_size,
sector_t cache_block_size)
{
unsigned i;
unsigned nr_sentinels_per_queue = 2u * NR_CACHE_LEVELS;
unsigned total_sentinels = 2u * nr_sentinels_per_queue;
struct smq_policy *mq = kzalloc(sizeof(*mq), GFP_KERNEL);
if (!mq)
return NULL;
init_policy_functions(mq);
mq->cache_size = cache_size;
mq->cache_block_size = cache_block_size;
calc_hotspot_params(origin_size, cache_block_size, from_cblock(cache_size),
&mq->hotspot_block_size, &mq->nr_hotspot_blocks);
mq->cache_blocks_per_hotspot_block = div64_u64(mq->hotspot_block_size, mq->cache_block_size);
mq->hotspot_level_jump = 1u;
if (space_init(&mq->es, total_sentinels + mq->nr_hotspot_blocks + from_cblock(cache_size))) {
DMERR("couldn't initialize entry space");
goto bad_pool_init;
}
init_allocator(&mq->writeback_sentinel_alloc, &mq->es, 0, nr_sentinels_per_queue);
for (i = 0; i < nr_sentinels_per_queue; i++)
get_entry(&mq->writeback_sentinel_alloc, i)->sentinel = true;
init_allocator(&mq->demote_sentinel_alloc, &mq->es, nr_sentinels_per_queue, total_sentinels);
for (i = 0; i < nr_sentinels_per_queue; i++)
get_entry(&mq->demote_sentinel_alloc, i)->sentinel = true;
init_allocator(&mq->hotspot_alloc, &mq->es, total_sentinels,
total_sentinels + mq->nr_hotspot_blocks);
init_allocator(&mq->cache_alloc, &mq->es,
total_sentinels + mq->nr_hotspot_blocks,
total_sentinels + mq->nr_hotspot_blocks + from_cblock(cache_size));
mq->hotspot_hit_bits = alloc_bitset(mq->nr_hotspot_blocks);
if (!mq->hotspot_hit_bits) {
DMERR("couldn't allocate hotspot hit bitset");
goto bad_hotspot_hit_bits;
}
clear_bitset(mq->hotspot_hit_bits, mq->nr_hotspot_blocks);
if (from_cblock(cache_size)) {
mq->cache_hit_bits = alloc_bitset(from_cblock(cache_size));
if (!mq->cache_hit_bits) {
dm cache: add stochastic-multi-queue (smq) policy The stochastic-multi-queue (smq) policy addresses some of the problems with the current multiqueue (mq) policy. Memory usage ------------ The mq policy uses a lot of memory; 88 bytes per cache block on a 64 bit machine. SMQ uses 28bit indexes to implement it's data structures rather than pointers. It avoids storing an explicit hit count for each block. It has a 'hotspot' queue rather than a pre cache which uses a quarter of the entries (each hotspot block covers a larger area than a single cache block). All these mean smq uses ~25bytes per cache block. Still a lot of memory, but a substantial improvement nontheless. Level balancing --------------- MQ places entries in different levels of the multiqueue structures based on their hit count (~ln(hit count)). This means the bottom levels generally have the most entries, and the top ones have very few. Having unbalanced levels like this reduces the efficacy of the multiqueue. SMQ does not maintain a hit count, instead it swaps hit entries with the least recently used entry from the level above. The over all ordering being a side effect of this stochastic process. With this scheme we can decide how many entries occupy each multiqueue level, resulting in better promotion/demotion decisions. Adaptability ------------ The MQ policy maintains a hit count for each cache block. For a different block to get promoted to the cache it's hit count has to exceed the lowest currently in the cache. This means it can take a long time for the cache to adapt between varying IO patterns. Periodically degrading the hit counts could help with this, but I haven't found a nice general solution. SMQ doesn't maintain hit counts, so a lot of this problem just goes away. In addition it tracks performance of the hotspot queue, which is used to decide which blocks to promote. If the hotspot queue is performing badly then it starts moving entries more quickly between levels. This lets it adapt to new IO patterns very quickly. Performance ----------- In my tests SMQ shows substantially better performance than MQ. Once this matures a bit more I'm sure it'll become the default policy. Signed-off-by: Joe Thornber <ejt@redhat.com> Signed-off-by: Mike Snitzer <snitzer@redhat.com>
2015-05-15 14:33:34 +00:00
DMERR("couldn't allocate cache hit bitset");
goto bad_cache_hit_bits;
}
clear_bitset(mq->cache_hit_bits, from_cblock(mq->cache_size));
} else
mq->cache_hit_bits = NULL;
mq->tick = 0;
spin_lock_init(&mq->lock);
dm cache: add stochastic-multi-queue (smq) policy The stochastic-multi-queue (smq) policy addresses some of the problems with the current multiqueue (mq) policy. Memory usage ------------ The mq policy uses a lot of memory; 88 bytes per cache block on a 64 bit machine. SMQ uses 28bit indexes to implement it's data structures rather than pointers. It avoids storing an explicit hit count for each block. It has a 'hotspot' queue rather than a pre cache which uses a quarter of the entries (each hotspot block covers a larger area than a single cache block). All these mean smq uses ~25bytes per cache block. Still a lot of memory, but a substantial improvement nontheless. Level balancing --------------- MQ places entries in different levels of the multiqueue structures based on their hit count (~ln(hit count)). This means the bottom levels generally have the most entries, and the top ones have very few. Having unbalanced levels like this reduces the efficacy of the multiqueue. SMQ does not maintain a hit count, instead it swaps hit entries with the least recently used entry from the level above. The over all ordering being a side effect of this stochastic process. With this scheme we can decide how many entries occupy each multiqueue level, resulting in better promotion/demotion decisions. Adaptability ------------ The MQ policy maintains a hit count for each cache block. For a different block to get promoted to the cache it's hit count has to exceed the lowest currently in the cache. This means it can take a long time for the cache to adapt between varying IO patterns. Periodically degrading the hit counts could help with this, but I haven't found a nice general solution. SMQ doesn't maintain hit counts, so a lot of this problem just goes away. In addition it tracks performance of the hotspot queue, which is used to decide which blocks to promote. If the hotspot queue is performing badly then it starts moving entries more quickly between levels. This lets it adapt to new IO patterns very quickly. Performance ----------- In my tests SMQ shows substantially better performance than MQ. Once this matures a bit more I'm sure it'll become the default policy. Signed-off-by: Joe Thornber <ejt@redhat.com> Signed-off-by: Mike Snitzer <snitzer@redhat.com>
2015-05-15 14:33:34 +00:00
q_init(&mq->hotspot, &mq->es, NR_HOTSPOT_LEVELS);
mq->hotspot.nr_top_levels = 8;
mq->hotspot.nr_in_top_levels = min(mq->nr_hotspot_blocks / NR_HOTSPOT_LEVELS,
from_cblock(mq->cache_size) / mq->cache_blocks_per_hotspot_block);
q_init(&mq->clean, &mq->es, NR_CACHE_LEVELS);
q_init(&mq->dirty, &mq->es, NR_CACHE_LEVELS);
stats_init(&mq->hotspot_stats, NR_HOTSPOT_LEVELS);
stats_init(&mq->cache_stats, NR_CACHE_LEVELS);
if (h_init(&mq->table, &mq->es, from_cblock(cache_size)))
goto bad_alloc_table;
if (h_init(&mq->hotspot_table, &mq->es, mq->nr_hotspot_blocks))
goto bad_alloc_hotspot_table;
sentinels_init(mq);
mq->write_promote_level = mq->read_promote_level = NR_HOTSPOT_LEVELS;
mq->next_hotspot_period = jiffies;
mq->next_cache_period = jiffies;
return &mq->policy;
bad_alloc_hotspot_table:
h_exit(&mq->table);
bad_alloc_table:
free_bitset(mq->cache_hit_bits);
bad_cache_hit_bits:
free_bitset(mq->hotspot_hit_bits);
bad_hotspot_hit_bits:
space_exit(&mq->es);
bad_pool_init:
kfree(mq);
return NULL;
}
/*----------------------------------------------------------------*/
static struct dm_cache_policy_type smq_policy_type = {
.name = "smq",
.version = {1, 0, 0},
.hint_size = 4,
.owner = THIS_MODULE,
.create = smq_create
};
static struct dm_cache_policy_type default_policy_type = {
.name = "default",
.version = {1, 4, 0},
.hint_size = 4,
.owner = THIS_MODULE,
.create = smq_create,
.real = &smq_policy_type
};
dm cache: add stochastic-multi-queue (smq) policy The stochastic-multi-queue (smq) policy addresses some of the problems with the current multiqueue (mq) policy. Memory usage ------------ The mq policy uses a lot of memory; 88 bytes per cache block on a 64 bit machine. SMQ uses 28bit indexes to implement it's data structures rather than pointers. It avoids storing an explicit hit count for each block. It has a 'hotspot' queue rather than a pre cache which uses a quarter of the entries (each hotspot block covers a larger area than a single cache block). All these mean smq uses ~25bytes per cache block. Still a lot of memory, but a substantial improvement nontheless. Level balancing --------------- MQ places entries in different levels of the multiqueue structures based on their hit count (~ln(hit count)). This means the bottom levels generally have the most entries, and the top ones have very few. Having unbalanced levels like this reduces the efficacy of the multiqueue. SMQ does not maintain a hit count, instead it swaps hit entries with the least recently used entry from the level above. The over all ordering being a side effect of this stochastic process. With this scheme we can decide how many entries occupy each multiqueue level, resulting in better promotion/demotion decisions. Adaptability ------------ The MQ policy maintains a hit count for each cache block. For a different block to get promoted to the cache it's hit count has to exceed the lowest currently in the cache. This means it can take a long time for the cache to adapt between varying IO patterns. Periodically degrading the hit counts could help with this, but I haven't found a nice general solution. SMQ doesn't maintain hit counts, so a lot of this problem just goes away. In addition it tracks performance of the hotspot queue, which is used to decide which blocks to promote. If the hotspot queue is performing badly then it starts moving entries more quickly between levels. This lets it adapt to new IO patterns very quickly. Performance ----------- In my tests SMQ shows substantially better performance than MQ. Once this matures a bit more I'm sure it'll become the default policy. Signed-off-by: Joe Thornber <ejt@redhat.com> Signed-off-by: Mike Snitzer <snitzer@redhat.com>
2015-05-15 14:33:34 +00:00
static int __init smq_init(void)
{
int r;
r = dm_cache_policy_register(&smq_policy_type);
if (r) {
DMERR("register failed %d", r);
return -ENOMEM;
}
r = dm_cache_policy_register(&default_policy_type);
if (r) {
DMERR("register failed (as default) %d", r);
dm_cache_policy_unregister(&smq_policy_type);
return -ENOMEM;
}
dm cache: add stochastic-multi-queue (smq) policy The stochastic-multi-queue (smq) policy addresses some of the problems with the current multiqueue (mq) policy. Memory usage ------------ The mq policy uses a lot of memory; 88 bytes per cache block on a 64 bit machine. SMQ uses 28bit indexes to implement it's data structures rather than pointers. It avoids storing an explicit hit count for each block. It has a 'hotspot' queue rather than a pre cache which uses a quarter of the entries (each hotspot block covers a larger area than a single cache block). All these mean smq uses ~25bytes per cache block. Still a lot of memory, but a substantial improvement nontheless. Level balancing --------------- MQ places entries in different levels of the multiqueue structures based on their hit count (~ln(hit count)). This means the bottom levels generally have the most entries, and the top ones have very few. Having unbalanced levels like this reduces the efficacy of the multiqueue. SMQ does not maintain a hit count, instead it swaps hit entries with the least recently used entry from the level above. The over all ordering being a side effect of this stochastic process. With this scheme we can decide how many entries occupy each multiqueue level, resulting in better promotion/demotion decisions. Adaptability ------------ The MQ policy maintains a hit count for each cache block. For a different block to get promoted to the cache it's hit count has to exceed the lowest currently in the cache. This means it can take a long time for the cache to adapt between varying IO patterns. Periodically degrading the hit counts could help with this, but I haven't found a nice general solution. SMQ doesn't maintain hit counts, so a lot of this problem just goes away. In addition it tracks performance of the hotspot queue, which is used to decide which blocks to promote. If the hotspot queue is performing badly then it starts moving entries more quickly between levels. This lets it adapt to new IO patterns very quickly. Performance ----------- In my tests SMQ shows substantially better performance than MQ. Once this matures a bit more I'm sure it'll become the default policy. Signed-off-by: Joe Thornber <ejt@redhat.com> Signed-off-by: Mike Snitzer <snitzer@redhat.com>
2015-05-15 14:33:34 +00:00
return 0;
}
static void __exit smq_exit(void)
{
dm_cache_policy_unregister(&smq_policy_type);
dm_cache_policy_unregister(&default_policy_type);
dm cache: add stochastic-multi-queue (smq) policy The stochastic-multi-queue (smq) policy addresses some of the problems with the current multiqueue (mq) policy. Memory usage ------------ The mq policy uses a lot of memory; 88 bytes per cache block on a 64 bit machine. SMQ uses 28bit indexes to implement it's data structures rather than pointers. It avoids storing an explicit hit count for each block. It has a 'hotspot' queue rather than a pre cache which uses a quarter of the entries (each hotspot block covers a larger area than a single cache block). All these mean smq uses ~25bytes per cache block. Still a lot of memory, but a substantial improvement nontheless. Level balancing --------------- MQ places entries in different levels of the multiqueue structures based on their hit count (~ln(hit count)). This means the bottom levels generally have the most entries, and the top ones have very few. Having unbalanced levels like this reduces the efficacy of the multiqueue. SMQ does not maintain a hit count, instead it swaps hit entries with the least recently used entry from the level above. The over all ordering being a side effect of this stochastic process. With this scheme we can decide how many entries occupy each multiqueue level, resulting in better promotion/demotion decisions. Adaptability ------------ The MQ policy maintains a hit count for each cache block. For a different block to get promoted to the cache it's hit count has to exceed the lowest currently in the cache. This means it can take a long time for the cache to adapt between varying IO patterns. Periodically degrading the hit counts could help with this, but I haven't found a nice general solution. SMQ doesn't maintain hit counts, so a lot of this problem just goes away. In addition it tracks performance of the hotspot queue, which is used to decide which blocks to promote. If the hotspot queue is performing badly then it starts moving entries more quickly between levels. This lets it adapt to new IO patterns very quickly. Performance ----------- In my tests SMQ shows substantially better performance than MQ. Once this matures a bit more I'm sure it'll become the default policy. Signed-off-by: Joe Thornber <ejt@redhat.com> Signed-off-by: Mike Snitzer <snitzer@redhat.com>
2015-05-15 14:33:34 +00:00
}
module_init(smq_init);
module_exit(smq_exit);
MODULE_AUTHOR("Joe Thornber <dm-devel@redhat.com>");
MODULE_LICENSE("GPL");
MODULE_DESCRIPTION("smq cache policy");
MODULE_ALIAS("dm-cache-default");