On 2024-08-12 21:54, Stanislav Fomichev wrote:
On 08/12, Martin Karsten wrote:
On 2024-08-12 19:03, Stanislav Fomichev wrote:
On 08/12, Martin Karsten wrote:
On 2024-08-12 16:19, Stanislav Fomichev wrote:
On 08/12, Joe Damato wrote:
Greetings:
Martin Karsten (CC'd) and I have been collaborating on some ideas about
ways of reducing tail latency when using epoll-based busy poll and we'd
love to get feedback from the list on the code in this series. This is
the idea I mentioned at netdev conf, for those who were there. Barring
any major issues, we hope to submit this officially shortly after RFC.
The basic idea for suspending IRQs in this manner was described in an
earlier paper presented at Sigmetrics 2024 [1].
Let me explicitly call out the paper. Very nice analysis!
Thank you!
[snip]
Here's how it is intended to work:
- An administrator sets the existing sysfs parameters for
defer_hard_irqs and gro_flush_timeout to enable IRQ deferral.
- An administrator sets the new sysfs parameter irq_suspend_timeout
to a larger value than gro-timeout to enable IRQ suspension.
Can you expand more on what's the problem with the existing gro_flush_timeout?
Is it defer_hard_irqs_count? Or you want a separate timeout only for the
perfer_busy_poll case(why?)? Because looking at the first two patches,
you essentially replace all usages of gro_flush_timeout with a new variable
and I don't see how it helps.
gro-flush-timeout (in combination with defer-hard-irqs) is the default irq
deferral mechanism and as such, always active when configured. Its static
periodic softirq processing leads to a situation where:
- A long gro-flush-timeout causes high latencies when load is sufficiently
below capacity, or
- a short gro-flush-timeout causes overhead when softirq execution
asynchronously competes with application processing at high load.
The shortcomings of this are documented (to some extent) by our experiments.
See defer20 working well at low load, but having problems at high load,
while defer200 having higher latency at low load.
irq-suspend-timeout is only active when an application uses
prefer-busy-polling and in that case, produces a nice alternating pattern of
application processing and networking processing (similar to what we
describe in the paper). This then works well with both low and high load.
So you only want it for the prefer-busy-pollingc case, makes sense. I was
a bit confused by the difference between defer200 and suspend200,
but now I see that defer200 does not enable busypoll.
I'm assuming that if you enable busypool in defer200 case, the numbers
should be similar to suspend200 (ignoring potentially affecting
non-busypolling queues due to higher gro_flush_timeout).
defer200 + napi busy poll is essentially what we labelled "busy" and it does
not perform as well, since it still suffers interference between application
and softirq processing.
With all your patches applied? Why? Userspace not keeping up?
Note our "busy" case does not utilize our patches.
As illustrated by our performance numbers, its performance is better
than the base case, but at the cost of higher cpu utilization and it's
still not as good as suspend20.
Explanation (conjecture):
It boils down to having to set a particular static value for
gro-flush-timeout that is then always active.
If busy-poll + application processing takes longer than this timeout,
the next softirq runs while the application is still active, which
causes interference.
Once a softirq runs, the irq-loop (Loop 2) takes control. When the
application thread comes back to epoll_wait, it already finds data, thus
ep_poll does not run napi_busy_poll at all, thus the irq-loop stays in
control.
This continues until by chance the application finds no readily
available data when calling epoll_wait and ep_poll runs another
napi_busy_poll. Then the system switches back to busy-polling mode.
So essentially the system non-deterministically alternates between
busy-polling and irq deferral. irq deferral determines the high-order
tail latencies, but there is still enough interference to make a
difference. It's not as bad as in the base case, but not as good as
properly controlled irq suspension.
Maybe expand more on what code paths are we trying to improve? Existing
busy polling code is not super readable, so would be nice to simplify
it a bit in the process (if possible) instead of adding one more tunable.
There are essentially three possible loops for network processing:
1) hardirq -> softirq -> napi poll; this is the baseline functionality
2) timer -> softirq -> napi poll; this is deferred irq processing scheme
with the shortcomings described above
3) epoll -> busy-poll -> napi poll
If a system is configured for 1), not much can be done, as it is difficult
to interject anything into this loop without adding state and side effects.
This is what we tried for the paper, but it ended up being a hack.
If however the system is configured for irq deferral, Loops 2) and 3)
"wrestle" with each other for control. Injecting the larger
irq-suspend-timeout for 'timer' in Loop 2) essentially tilts this in favour
of Loop 3) and creates the nice pattern describe above.
And you hit (2) when the epoll goes to sleep and/or when the userspace
isn't fast enough to keep up with the timer, presumably? I wonder
if need to use this opportunity and do proper API as Joe hints in the
cover letter. Something over netlink to say "I'm gonna busy-poll on
this queue / napi_id and with this timeout". And then we can essentially make
gro_flush_timeout per queue (and avoid
napi_resume_irqs/napi_suspend_irqs). Existing gro_flush_timeout feels
too hacky already :-(
If someone would implement the necessary changes to make these parameters
per-napi, this would improve things further, but note that the current
proposal gives strong performance across a range of workloads, which is
otherwise difficult to impossible to achieve.
Let's see what other people have to say. But we tried to do a similar
setup at Google recently and getting all these parameters right
was not trivial. Joe's recent patch series to push some of these into
epoll context are a step in the right direction. It would be nice to
have more explicit interface to express busy poling preference for
the users vs chasing a bunch of global tunables and fighting against softirq
wakups.
One of the goals of this patch set is to reduce parameter tuning and
make the parameter setting independent of workload dynamics, so it
should make things easier. This is of course notwithstanding that
per-napi settings would be even better.
If you are able to share more details of your previous experiments (here
or off-list), I would be very interested.
Note that napi_suspend_irqs/napi_resume_irqs is needed even for the sake of
an individual queue or application to make sure that IRQ suspension is
enabled/disabled right away when the state of the system changes from busy
to idle and back.
Can we not handle everything in napi_busy_loop? If we can mark some napi
contexts as "explicitly polled by userspace with a larger defer timeout",
we should be able to do better compared to current NAPI_F_PREFER_BUSY_POLL
which is more like "this particular napi_poll call is user busy polling".
Then either the application needs to be polling all the time (wasting
cpu cycles) or latencies will be determined by the timeout.
Only when switching back and forth between polling and interrupts is it
possible to get low latencies across a large spectrum of offered loads
without burning cpu cycles at 100%.
[snip]
- suspendX:
- set defer_hard_irqs to 100
- set gro_flush_timeout to X,000
- set irq_suspend_timeout to 20,000,000
- enable busy poll via the existing ioctl (busy_poll_usecs = 0,
busy_poll_budget = 64, prefer_busy_poll = true)
What's the intention of `busy_poll_usecs = 0` here? Presumably we fallback
to busy_poll sysctl value?
Before this patch set, ep_poll only calls napi_busy_poll, if busy_poll
(sysctl) or busy_poll_usecs is nonzero. However, this might lead to
busy-polling even when the application does not actually need or want it.
Only one iteration through the busy loop is needed to make the new scheme
work. Additional napi busy polling over and above is optional.
Ack, thanks, was trying to understand why not stay with
busy_poll_usecs=64 for consistency. But I guess you were just
trying to show that patch 4/5 works.
Right, and we would potentially be wasting CPU cycles by adding more
busy-looping.
Or potentially improving the latency more if you happen to get more packets
during busy_poll_usecs duration? I'd imagine some applications might
prefer to 100% busy poll without ever going to sleep (that would probably
require getting rid of napi_id tracking in epoll, but that's a different story).
Yes, one could do full application-to-napi busy polling. The performance
would be slightly better than irq suspension, but it would be quite
wasteful during low load. One premise for our work is that saving cycles
is a meaningful objective.
Thanks,
Martin
