Kirr: In particular with
ethtool -C <ifname> rx-usecs 0 rx-frames 0
now it is possible to disable RX delays when NIC usage requires low-latency.
See this thread for context:
https://www.spinics.net/lists/netdev/msg217665.html
My specific case is that:
We have many computers with gigabit Realtek NICs. For 2 such computers
connected to a gigabit store-and-forward switch the minimum round-trip
time for small pings (`ping -i 0 -w 3 -s 56 -q peer`) is ~ 30μs.
However it turned out that when Ethernet frame length transitions 127 ->
128 bytes (`ping -i 0 -w 3 -s {81 -> 82} -q peer`) the lowest RTT
transitions step-wise to ~ 270μs.
As David Light said this is RX interrupt mitigation done by NIC which creates
the latency. For workloads when low-latency is required with e.g. Intel,
BCM etc NIC drivers one just uses `ethtool -C rx-usecs ...` to reduce
the time NIC delays before interrupting CPU, but it turned out
`ethtool -C` is not supported by r8169 driver.
Like Stéphane ANCELOT I've traced the problem down to IntrMitigate being
hardcoded to != 0 for our chips (we have 8168 based NICs):
https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/tree/drivers/net/ethernet/realtek/r8169.c#n5460
static void rtl_hw_start_8169(struct net_device *dev) {
...
/*
* Undocumented corner. Supposedly:
* (TxTimer << 12) | (TxPackets << 8) | (RxTimer << 4) | RxPackets
*/
RTL_W16(IntrMitigate, 0x0000);
https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/tree/drivers/net/ethernet/realtek/r8169.c#n6346
static void rtl_hw_start_8168(struct net_device *dev) {
...
RTL_W16(IntrMitigate, 0x5151);
and then I've also found
https://www.spinics.net/lists/netdev/msg217665.html
and original Francois' patch:
https://www.spinics.net/lists/netdev/msg217984.htmlhttps://www.spinics.net/lists/netdev/msg218207.html
So could we please finally get support for tuning r8169 interrupt
coalescing in tree? (so that next poor soul who hits the problem does
not need to go all the way to dig into driver sources and internet
wildly and finally patch locally
-RTL_W16(IntrMitigate, 0x5151);
+RTL_W16(IntrMitigate, 0x5100);
guessing whether it is right or not and also having to care to deploy
the patch everywhere it needs to be used, etc...).
To do so I've took original Francois's patch from 2012 and reworked it a bit:
- updated to latest net-next.git;
- adjusted scaling setup based on feedback from Hayes to pick up scaling
vector depending not only on link speed but also on CPlusCmd[0:1] and to
adjust CPlusCmd[0:1] correspondingly when setting timings;
- improved a bit (I think so) error handling.
I've tested the patch on "RTL8168d/8111d" (XID 083000c0) and with it and
`ethtool -C rx-usecs 0 rx-frames 0` on both ends it improves:
- minimum RTT latency:
~270μs -> ~30μs (small packet),
~330μs -> ~110μs (full 1.5K ethernet frame)
- average RTT latency:
~480μs -> ~50μs (small packet),
~560μs -> ~125μs (full 1.5K ethernet frame)
( before:
root@neo1:# ping -i 0 -w 3 -s 82 -q neo2
PING neo2.kirr.nexedi.com (192.168.102.21) 82(110) bytes of data.
--- neo2.kirr.nexedi.com ping statistics ---
5906 packets transmitted, 5905 received, 0% packet loss, time 2999ms
rtt min/avg/max/mdev = 0.274/0.485/0.607/0.026 ms, ipg/ewma 0.508/0.489 ms
root@neo1:# ping -i 0 -w 3 -s 1472 -q neo2
PING neo2.kirr.nexedi.com (192.168.102.21) 1472(1500) bytes of data.
--- neo2.kirr.nexedi.com ping statistics ---
5073 packets transmitted, 5073 received, 0% packet loss, time 2999ms
rtt min/avg/max/mdev = 0.330/0.566/0.710/0.028 ms, ipg/ewma 0.591/0.544 ms
after:
root@neo1# ping -i 0 -w 3 -s 82 -q neo2
PING neo2.kirr.nexedi.com (192.168.102.21) 82(110) bytes of data.
--- neo2.kirr.nexedi.com ping statistics ---
45815 packets transmitted, 45815 received, 0% packet loss, time 3000ms
rtt min/avg/max/mdev = 0.036/0.051/0.368/0.010 ms, ipg/ewma 0.065/0.053 ms
root@neo1:# ping -i 0 -w 3 -s 1472 -q neo2
PING neo2.kirr.nexedi.com (192.168.102.21) 1472(1500) bytes of data.
--- neo2.kirr.nexedi.com ping statistics ---
21250 packets transmitted, 21250 received, 0% packet loss, time 3000ms
rtt min/avg/max/mdev = 0.112/0.125/0.390/0.007 ms, ipg/ewma 0.141/0.125 ms
the small -> 1.5K latency growth is understandable as it takes ~15μs
to transmit 1.5K on 1Gbps on the wire and with 2 hosts and 1 switch
and ICMP ECHO + ECHO reply the packet has to travel 4 ethernet
segments which is already 60μs;
probably something a bit else is also there as e.g. on Linux, even
with `cpupower frequency-set -g performance`, on some computers I've
noticed the kernel can be spending more time in software-only mode
when incoming packets go in less frequently. E.g. this program can
demonstrate the effect for ICMP ECHO processing:
https://lab.nexedi.com/kirr/bcc/blob/43cfc13b/tools/pinglat.py
(later this was found to be partly due to C-states exit latencies) )
We have this patch running in our testing setup for 1 months already
without any issues observed.
It remains to be clarified whether RX and TX timers use the same base.
For now I've set them equally, but Francois's original patch version
suggests it could be not the same.
I've got no feedback at all to my original posting of this patch and questions
https://www.spinics.net/lists/netdev/msg457173.html
neither from Francois, nor from any people from Realtek during one month.
So I suggest we simply apply it to net-next.git now.
Cc: Francois Romieu <romieu@fr.zoreil.com>
Cc: Hayes Wang <hayeswang@realtek.com>
Cc: Realtek linux nic maintainers <nic_swsd@realtek.com>
Cc: David Laight <David.Laight@ACULAB.COM>
Cc: Stéphane ANCELOT <sancelot@free.fr>
Cc: Eric Dumazet <edumazet@google.com>
Signed-off-by: Kirill Smelkov <kirr@nexedi.com>
Tested-by: Holger Hoffstätte <holger@applied-asynchrony.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
509708310c