Re: [PATCH RESEND 8/8] af_unix: charge buffers to kmemcg

[Sudeep Holla]

On 23/08/16 17:44, Vladimir Davydov wrote:
> Hello,
>
> On Tue, Aug 23, 2016 at 02:48:11PM +0100, Sudeep K N wrote:
> > On Tue, May 24, 2016 at 5:36 PM, Vladimir Davydov
> > <vdavydov@xxxxxxxxxxxxx> wrote:
> > > On Tue, May 24, 2016 at 06:02:06AM -0700, Eric Dumazet wrote:
> > > > On Tue, 2016-05-24 at 11:49 +0300, Vladimir Davydov wrote:
> > > > > Unix sockets can consume a significant amount of system memory, hence
> > > > > they should be accounted to kmemcg.
> > > > >
> > > > > Since unix socket buffers are always allocated from process context,
> > > > > all we need to do to charge them to kmemcg is set __GFP_ACCOUNT in
> > > > > sock->sk_allocation mask.
> > > >
> > > > I have two questions :
> > > >
> > > > 1) What happens when a buffer, allocated from socket <A> lands in a
> > > > different socket <B>, maybe owned by another user/process.
> > > >
> > > > Who owns it now, in term of kmemcg accounting ?
> > >
> > > We never move memcg charges. E.g. if two processes from different
> > > cgroups are sharing a memory region, each page will be charged to the
> > > process which touched it first. Or if two processes are working with the
> > > same directory tree, inodes and dentries will be charged to the first
> > > user. The same is fair for unix socket buffers - they will be charged to
> > > the sender.
> > >
> > > > 2) Has performance impact been evaluated ?
> > >
> > > I ran netperf STREAM_STREAM with default options in a kmemcg on
> > > a 4 core x 2 HT box. The results are below:
> > >
> > >  # clients            bandwidth (10^6bits/sec)
> > >                     base              patched
> > >          1      67643 +-  725      64874 +-  353    - 4.0 %
> > >          4     193585 +- 2516     186715 +- 1460    - 3.5 %
> > >          8     194820 +-  377     187443 +- 1229    - 3.7 %
> > >
> > > So the accounting doesn't come for free - it takes ~4% of performance.
> > > I believe we could optimize it by using per cpu batching not only on
> > > charge, but also on uncharge in memcg core, but that's beyond the scope
> > > of this patch set - I'll take a look at this later.
> > >
> > > Anyway, if performance impact is found to be unacceptable, it is always
> > > possible to disable kmem accounting at boot time (cgroup.memory=nokmem)
> > > or not use memory cgroups at runtime at all (thanks to jump labels
> > > there'll be no overhead even if they are compiled in).
> >
> > I started seeing almost 10% degradation in the hackbench score with v4.8-rc1
> > Bisecting it resulted in this patch, i.e. Commit 3aa9799e1364 ("af_unix: charge
> > buffers to kmemcg") in the mainline.
> >
> > As per the commit log, it seems like that's expected but I was not sure about
> > the margin. I also see the hackbench score is more inconsistent after this
> > patch, but I may be wrong as that's based on limited observation.
> >
> > Is this something we can ignore as hackbench is more synthetic compared
> > to the gain this patch provides in some real workloads ?
>
> AFAIU hackbench essentially measures the rate of sending data over a
> unix socket back and forth between processes running on different cpus,
> so it isn't a surprise that the patch resulted in a degradation, as it
> makes every skb page allocation/deallocation inc/dec an atomic counter
> inside memcg. The more processes/cpus running in the same cgroup are
> involved in this test, the more significant the overhead of this atomic
> counter is going to be.

Understood.

> The degradation is not unavoidable - it can be fixed by making kmem
> charge/uncharge code use per-cpu batches. The infrastructure for this
> already exists in memcontrol.c. If it were not for the legacy
> mem_cgroup->kmem counter (which is actually useless and will be dropped
> in cgroup v2), the issue would be pretty easy to fix. However, this
> legacy counter makes a possible implementation quite messy, so I'd like
> to postpone it until cgroup v2 has finally settled down.

Sure

> Regarding your problem. As a workaround you can either start your
> workload in the root memory cgroup or disable kmem accounting for memory
> cgroups altogether (via cgroup.memory=nokmem boot option). If you find
> the issue critical, I don't mind reverting the patch - we can always
> re-apply it once per-cpu batches are implemented for kmem charges.

I did try "cgroup.memory=nokmem" as specified in the commit message, I
saw the result to be not so consistent. I need to check again to be sure.

I am not asking to revert, just wanted to know if that's expected so
that we can adjust the scores when comparing especially if we are using
it as some kind of benchmark in development.

--
Regards,
Sudeep