Skip to contentSkip to conversation

sched: prefer an idle cpu vs an idle sibling for BALANCE_WAKE

PATCH Josef Bacik 3 participants 4 messages May 27, 2015 – May 28, 2015 1 message not in the archive

lore
Share & tools
lore

A message that is not in the archive (sent off-list or not yet received): [email protected]

#1 lore raw
Quoted 8 lines · Josef Bacik

At Facebook we have a pretty heavily multi-threaded application that is sensitive to latency. We have been pulling forward the old SD_WAKE_IDLE code because it gives us a pretty significant performance gain (like 20%). It turns out this is because there are cases where the scheduler puts our task on a busy CPU when there are idle CPU's in the system. We verify this by reading the cpu_delay_req_avg_us from the scheduler netlink stuff. With our crappy patch we get much lower numbers vs baseline.

Quoted 8 lines

SD_BALANCE_WAKE is supposed to find us an idle cpu to run on, however it is just looking for an idle sibling, preferring affinity over all else. This is not helpful in all cases, and SD_BALANCE_WAKE's job is to find us an idle cpu, not garuntee affinity. Fix this by first trying to find an idle sibling, and then if the cpu is not idle fall through to the logic to find an idle cpu. With this patch we get slightly better performance than with our forward port of SD_WAKE_IDLE. Thanks,

I rigged up a test script to run the perf bench sched tests and give me the numbers. Here are the numbers

4.0

Messaging: 56.934 Total runtime in seconds
Pipe: 105620.762 ops/sec

4.0 + my patch

Messaging: 47.374
Pipe: 113691.199

so ~20% better performance out of the Messaging test which is sort of like HHVM and ~8% better pipe performance. This box is a 2 socket 16 core box. I've attached the script I'm using, basically I just run each thing 5 times, and for the perf bench sched pipe run I do NR_CPUS/2 instances of them in parallel.

If you are interested I'd be happy to show you numbers for our HHVM test, but they are less straightforward and require pretty pictures and a book of how to read the numbers. Thanks

Josef

Acked-by
#2 lore raw
Quoted 21 lines · Josef Bacik
Quoted 4 lines · Josef Bacik

SD_BALANCE_WAKE is supposed to find us an idle cpu to run on, however it is just
looking for an idle sibling, preferring affinity over all else. This is not

Quoted 9 lines

helpful in all cases, and SD_BALANCE_WAKE's job is to find us an idle cpu, not
garuntee affinity. Fix this by first trying to find an idle sibling, and then
if the cpu is not idle fall through to the logic to find an idle cpu. With this
patch we get slightly better performance than with our forward port of SD_WAKE_IDLE. Thanks,

I rigged up a test script to run the perf bench sched tests and give me the numbers. Here are the numbers

4.0

Quoted 7 lines

Messaging: 56.934 Total runtime in seconds
Pipe: 105620.762 ops/sec

4.0 + my patch

Messaging: 47.374
Pipe: 113691.199

I did not get the email with your original patch, either to my inbox or my lkml folder, but I saw the patch on pastebin, and it looks good.

When you resend it, please feel free to add my

Assuming the version you meant to email yesterday was the same one that you showed me on pastebin, of course :)

#3 lore raw
Quoted 34 lines · Rik van Riel
Quoted 21 lines · Josef Bacik
Quoted 4 lines · Josef Bacik

SD_BALANCE_WAKE is supposed to find us an idle cpu to run on, however it is just
looking for an idle sibling, preferring affinity over all else. This is not

Quoted 9 lines

helpful in all cases, and SD_BALANCE_WAKE's job is to find us an idle cpu, not
garuntee affinity. Fix this by first trying to find an idle sibling, and then
if the cpu is not idle fall through to the logic to find an idle cpu. With this
patch we get slightly better performance than with our forward port of SD_WAKE_IDLE. Thanks,

I rigged up a test script to run the perf bench sched tests and give me the numbers. Here are the numbers

4.0

Quoted 7 lines

Messaging: 56.934 Total runtime in seconds
Pipe: 105620.762 ops/sec

4.0 + my patch

Messaging: 47.374
Pipe: 113691.199

I did not get the email with your original patch, either to my inbox or my lkml folder, but I saw the patch on pastebin, and it looks good.

Quoted 7 lines

When you resend it, please feel free to add my

Assuming the version you meant to email yesterday was the same one that you showed me on pastebin, of course :)

Ha yes it's the same, sorry I'm not sure what happened, I've resent it again from a different machine, let me know if you don't get the new one and I'll just send it from thunderbird. Thanks,

Josef

#4 lore raw
In reply to Josef Bacik
Quoted 23 lines · Josef Bacik
Quoted 8 lines · Josef Bacik

At Facebook we have a pretty heavily multi-threaded application that is sensitive to latency. We have been pulling forward the old SD_WAKE_IDLE code because it gives us a pretty significant performance gain (like 20%). It turns out this is because there are cases where the scheduler puts our task on a busy CPU when there are idle CPU's in the system. We verify this by reading the cpu_delay_req_avg_us from the scheduler netlink stuff. With our crappy patch we get much lower numbers vs baseline.

Quoted 8 lines

SD_BALANCE_WAKE is supposed to find us an idle cpu to run on, however it is just looking for an idle sibling, preferring affinity over all else. This is not helpful in all cases, and SD_BALANCE_WAKE's job is to find us an idle cpu, not garuntee affinity. Fix this by first trying to find an idle sibling, and then if the cpu is not idle fall through to the logic to find an idle cpu. With this patch we get slightly better performance than with our forward port of SD_WAKE_IDLE. Thanks,

I rigged up a test script to run the perf bench sched tests and give me the numbers. Here are the numbers

4.0

Quoted 7 lines

Messaging: 56.934 Total runtime in seconds
Pipe: 105620.762 ops/sec

4.0 + my patch

Messaging: 47.374
Pipe: 113691.199

Btw., with perf bench you don't really need much extra scripting, something like this should give you pretty good numbers plus an stddev estimate:

   perf stat --null --repeat 10 perf bench sched messaging -l 10000

on my box this gives:

4.391469643 seconds time elapsed ( +- 2.81% )

you can adjust the -l value to move the runtime up/down to a value that you think runs long enough to give stable results.

Thanks,

Ingo