Date:   Thu, 18 Jun 2020 11:31:21 +0200
From:   Jesper Dangaard Brouer <brouer@...hat.com>
To:     Roman Gushchin <guro@...com>
Cc:     Vlastimil Babka <vbabka@...e.cz>,
        Shakeel Butt <shakeelb@...gle.com>,
        Andrew Morton <akpm@...ux-foundation.org>,
        Christoph Lameter <cl@...ux.com>,
        Johannes Weiner <hannes@...xchg.org>,
        Michal Hocko <mhocko@...nel.org>,
        Linux MM <linux-mm@...ck.org>,
        Kernel Team <kernel-team@...com>,
        LKML <linux-kernel@...r.kernel.org>,
        Mel Gorman <mgorman@...hsingularity.net>, brouer@...hat.com
Subject: Re: [PATCH v6 00/19] The new cgroup slab memory controller

On Thu, 18 Jun 2020 10:43:44 +0200
Jesper Dangaard Brouer <brouer@...hat.com> wrote:

> On Wed, 17 Jun 2020 18:29:28 -0700
> Roman Gushchin <guro@...com> wrote:
> 
> > On Wed, Jun 17, 2020 at 01:24:21PM +0200, Vlastimil Babka wrote:  
> > > On 6/17/20 5:32 AM, Roman Gushchin wrote:    
> > > > On Tue, Jun 16, 2020 at 08:05:39PM -0700, Shakeel Butt wrote:    
> > > >> On Tue, Jun 16, 2020 at 7:41 PM Roman Gushchin <guro@...com> wrote:    
> > > >> >
> > > >> > On Tue, Jun 16, 2020 at 06:46:56PM -0700, Shakeel Butt wrote:    
> > > >> > > On Mon, Jun 8, 2020 at 4:07 PM Roman Gushchin <guro@...com> wrote:    
> > > >> > > >    
> > > >> [...]    
> > > >> > >
> > > >> > > Have you performed any [perf] testing on SLAB with this patchset?    
> > > >> >
> > > >> > The accounting part is the same for SLAB and SLUB, so there should be no
> > > >> > significant difference. I've checked that it compiles, boots and passes
> > > >> > kselftests. And that memory savings are there.
> > > >> >    
> > > >> 
> > > >> What about performance? Also you mentioned that sharing kmem-cache
> > > >> between accounted and non-accounted can have additional overhead. Any
> > > >> difference between SLAB and SLUB for such a case?    
> > > > 
> > > > Not really.
> > > > 
> > > > Sharing a single set of caches adds some overhead to root- and non-accounted
> > > > allocations, which is something I've tried hard to avoid in my original version.
> > > > But I have to admit, it allows to simplify and remove a lot of code, and here
> > > > it's hard to argue with Johanness, who pushed on this design.
> > > > 
> > > > With performance testing it's not that easy, because it's not obvious what
> > > > we wanna test. Obviously, per-object accounting is more expensive, and
> > > > measuring something like 1000000 allocations and deallocations in a line from
> > > > a single kmem_cache will show a regression. But in the real world the relative
> > > > cost of allocations is usually low, and we can get some benefits from a smaller
> > > > working set and from having shared kmem_cache objects cache hot.
> > > > Not speaking about some extra memory and the fragmentation reduction.
> > > > 
> > > > We've done an extensive testing of the original version in Facebook production,
> > > > and we haven't noticed any regressions so far. But I have to admit, we were
> > > > using an original version with two sets of kmem_caches.
> > > > 
> > > > If you have any specific tests in mind, I can definitely run them. Or if you
> > > > can help with the performance evaluation, I'll appreciate it a lot.    
> > > 
> > > Jesper provided some pointers here [1], it would be really great if you could
> > > run at least those microbenchmarks. With mmtests it's the major question of
> > > which subset/profiles to run, maybe the referenced commits provide some hints,
> > > or maybe Mel could suggest what he used to evaluate SLAB vs SLUB not so long ago.
> > > 
> > > [1] https://lore.kernel.org/linux-mm/20200527103545.4348ac10@carbon/    
> > 
> > Oh, Jesper, I'm really sorry, somehow I missed your mail.
> > Thank you, Vlastimil, for pointing at it.
> > 
> > I've got some results (slab_bulk_test01), but honestly I fail to interpret them.
> > 
> > I ran original vs patched with SLUB and SLAB, each test several times and picked
> > 3 which looked most consistently. But it still looks very noisy.
> > 
> > I ran them on my desktop (8-core Ryzen 1700, 16 GB RAM, Fedora 32),
> > it's 5.8-rc1 + slab controller v6 vs 5.8-rc1 (default config from Fedora 32).  
> 
> What about running these tests on the server level hardware, that you
> intent to run this on?  

To give you an idea of the performance difference I ran the same test
on a Broadwell Intel(R) Xeon(R) CPU E5-1650 v4 @ 3.60GHz.

The SLUB fastpath:
 Type:kmem fastpath reuse Per elem: 60 cycles(tsc) 16.822 ns


> > 
> > How should I interpret this data?  
> 
> First of all these SLUB+SLAB microbenchmarks use object size 256 bytes,
> because network stack alloc object of this size for SKBs/sk_buff (due
> to cache-align as used size is 224 bytes). Checked SLUB: Each slab use
> 2 pages (8192 bytes) and contain 32 object of size 256 (256*32=8192).
> 
>   The SLUB allocator have a per-CPU slab which speedup fast-reuse, in this
> case up-to 32 objects. For SLUB the "fastpath reuse" test this behaviour,
> and it serves as a baseline for optimal 1-object performance (where my bulk
> API tries to beat that, which is possible even for 1-object due to knowing
> bulk API cannot be used from IRQ context).
> 
> SLUB fastpath: 3 measurements reporting cycles(tsc)
>  - SLUB-patched : fastpath reuse: 184 - 177 - 176  cycles(tsc)
>  - SLUB-original: fastpath reuse: 178 - 153 - 156  cycles(tsc)
> 

For your SLAB results:

 SLAB fastpath: 3 measurements reporting cycles(tsc)
  - SLAB-patched : 161 - 160 - 163  cycles(tsc)
  - SLAB-original: 174 - 170 - 159  cycles(tsc)

I find it strange that SLAB is slightly better than SLUB (in many
measurements), because SLUB should have an advantage on this fast-path
quick reuse due to the per-CPU slabs.  Maybe this is also related to
the CPU arch you are using?


> There are some stability concerns as you mention, but it seems pretty
> consistently that patched version is slower. If you compile with
> no-PREEMPT you can likely get more stable results (and remove a slight
> overhead for SLUB fastpath).
> 
> The microbenchmark also measures the bulk-API, which is AFAIK only used
> by network stack (and io_uring). I guess you shouldn't focus too much
> on these bulk measurements. When bulk-API cross this objects per slab
> threshold, or is unlucky is it use two per-CPU slab, then the
> measurements can fluctuate a bit.
> 
> Your numbers for SLUB bulk-API:
> 
> SLUB-patched - bulk-API
>  - SLUB-patched : bulk_quick_reuse objects=1 : 187 -  90 - 224  cycles(tsc)
>  - SLUB-patched : bulk_quick_reuse objects=2 : 110 -  53 - 133  cycles(tsc)
>  - SLUB-patched : bulk_quick_reuse objects=3 :  88 -  95 -  42  cycles(tsc)
>  - SLUB-patched : bulk_quick_reuse objects=4 :  91 -  85 -  36  cycles(tsc)
>  - SLUB-patched : bulk_quick_reuse objects=8 :  32 -  66 -  32  cycles(tsc)
> 
> SLUB-original -  bulk-API
>  - SLUB-original: bulk_quick_reuse objects=1 :  87 -  87 - 142  cycles(tsc)
>  - SLUB-original: bulk_quick_reuse objects=2 :  52 -  53 -  53  cycles(tsc)
>  - SLUB-original: bulk_quick_reuse objects=3 :  42 -  42 -  91  cycles(tsc)
>  - SLUB-original: bulk_quick_reuse objects=4 :  91 -  37 -  37  cycles(tsc)
>  - SLUB-original: bulk_quick_reuse objects=8 :  31 -  79 -  76  cycles(tsc)

Your numbers for SLAB bulk-API:

SLAB-patched -  bulk-API
 - SLAB-patched : bulk_quick_reuse objects=1 :  67 -  67 - 140  cycles(tsc)
 - SLAB-patched : bulk_quick_reuse objects=2 :  55 -  46 -  46  cycles(tsc)
 - SLAB-patched : bulk_quick_reuse objects=3 :  93 -  94 -  39  cycles(tsc)
 - SLAB-patched : bulk_quick_reuse objects=4 :  35 -  88 -  85  cycles(tsc)
 - SLAB-patched : bulk_quick_reuse objects=8 :  30 -  30 -  30  cycles(tsc)

SLAB-original-  bulk-API
 - SLAB-original: bulk_quick_reuse objects=1 : 143 - 136 -  67  cycles(tsc)
 - SLAB-original: bulk_quick_reuse objects=2 :  45 -  46 -  46  cycles(tsc)
 - SLAB-original: bulk_quick_reuse objects=3 :  38 -  39 -  39  cycles(tsc)
 - SLAB-original: bulk_quick_reuse objects=4 :  35 -  87 -  87  cycles(tsc)
 - SLAB-original: bulk_quick_reuse objects=8 :  29 -  66 -  30  cycles(tsc)

In case of SLAB I expect the bulk-API to be slightly faster than SLUB,
as the SLUB bulk code is much more advanced.



> Maybe it is just noise or instability in measurements, but it seem that the
> 1-object case is consistently slower in your patched version.
> 
> Mail is too long now... I'll take a look at your SLAB results and followup.

(This is my follow up with SLAB results.)

-- 
Best regards,
  Jesper Dangaard Brouer
  MSc.CS, Principal Kernel Engineer at Red Hat
  LinkedIn: http://www.linkedin.com/in/brouer

Powered by blists - more mailing lists