Re: [net] 4890b686f4: netperf.Throughput_Mbps -69.4% regression

[Oliver Sang <oliver.sang@intel.com>]

Hi all,

now we noticed this commit has already merged into mainline, and in our tests
there is still similar regression. [1]

not sure if there is a plan to merge some of the solutions that have been
discussed in this thread? we'll very glad to test patches if there is a request

Thanks a lot!

[1]
=========================================================================================
tbox_group/testcase/rootfs/kconfig/compiler/ip/runtime/nr_threads/cluster/send_size/test/cpufreq_governor/ucode:
  lkp-icl-2sp4/netperf/debian-11.1-x86_64-20220510.cgz/x86_64-rhel-8.3/gcc-11/ipv4/300s/50%/cs-localhost/10K/SCTP_STREAM_MANY/performance/0xd000363

7c80b038d23e1f4c 4890b686f4088c90432149bd6de
---------------- ---------------------------
         %stddev     %change         %stddev
             \          |                \
      9078           -55.9%       4006        netperf.Throughput_Mbps
    581006           -55.9%     256385        netperf.Throughput_total_Mbps
     36715           -54.6%      16674 ±  4%  netperf.time.involuntary_context_switches
      1885           -50.2%     938.33 ±  3%  netperf.time.percent_of_cpu_this_job_got
      5533           -49.9%       2771 ±  2%  netperf.time.system_time
    152.13           -59.5%      61.61 ±  2%  netperf.time.user_time
    418171 ±  5%     +89.4%     791954 ± 17%  netperf.time.voluntary_context_switches
 2.128e+09           -55.9%  9.389e+08        netperf.workload
     30217           +17.8%      35608        uptime.idle
 2.689e+10           +20.3%  3.234e+10        cpuidle..time
 6.366e+08           -48.1%  3.305e+08        cpuidle..usage
     70.26           +13.5       83.78        mpstat.cpu.all.idle%
      4.46            -1.5        2.92 ±  3%  mpstat.cpu.all.soft%
     23.71           -11.6       12.16 ±  3%  mpstat.cpu.all.sys%
      0.89            -0.5        0.38        mpstat.cpu.all.usr%
 1.392e+09           -57.5%   5.91e+08 ± 12%  numa-numastat.node0.local_node
 1.389e+09           -57.5%  5.906e+08 ± 12%  numa-numastat.node0.numa_hit
 1.369e+09           -54.5%  6.226e+08 ± 12%  numa-numastat.node1.local_node
 1.366e+09           -54.4%  6.222e+08 ± 12%  numa-numastat.node1.numa_hit
     36715           -54.6%      16674 ±  4%  time.involuntary_context_switches
      1885           -50.2%     938.33 ±  3%  time.percent_of_cpu_this_job_got
      5533           -49.9%       2771 ±  2%  time.system_time
    152.13           -59.5%      61.61 ±  2%  time.user_time
    418171 ±  5%     +89.4%     791954 ± 17%  time.voluntary_context_switches


On Tue, Jul 05, 2022 at 01:03:26PM +0800, Feng Tang wrote:
> On Sun, Jul 03, 2022 at 03:55:31PM -0700, Roman Gushchin wrote:
> > On Sun, Jul 03, 2022 at 06:43:53PM +0800, Feng Tang wrote:
> > > Hi Shakeel,
> > > 
> > > On Fri, Jul 01, 2022 at 08:47:29AM -0700, Shakeel Butt wrote:
> > > > On Mon, Jun 27, 2022 at 8:49 PM Feng Tang <feng.tang@intel.com> wrote:
> > > > > I just tested it, it does perform better (the 4th is with your patch),
> > > > > some perf-profile data is also listed.
> > > > >
> > > > >  7c80b038d23e1f4c 4890b686f4088c90432149bd6de 332b589c49656a45881bca4ecc0 e719635902654380b23ffce908d
> > > > > ---------------- --------------------------- --------------------------- ---------------------------
> > > > >      15722           -69.5%       4792           -40.8%       9300           -27.9%      11341        netperf.Throughput_Mbps
> > > > >
> > > > >       0.00            +0.3        0.26 ±  5%      +0.5        0.51            +1.3        1.27 ±  2%pp.self.__sk_mem_raise_allocated
> > > > >       0.00            +0.3        0.32 ± 15%      +1.7        1.74 ±  2%      +0.4        0.40 ±  2%  pp.self.propagate_protected_usage
> > > > >       0.00            +0.8        0.82 ±  7%      +0.9        0.90            +0.8        0.84        pp.self.__mod_memcg_state
> > > > >       0.00            +1.2        1.24 ±  4%      +1.0        1.01            +1.4        1.44        pp.self.try_charge_memcg
> > > > >       0.00            +2.1        2.06            +2.1        2.13            +2.1        2.11        pp.self.page_counter_uncharge
> > > > >       0.00            +2.1        2.14 ±  4%      +2.7        2.71            +2.6        2.60 ±  2%  pp.self.page_counter_try_charge
> > > > >       1.12 ±  4%      +3.1        4.24            +1.1        2.22            +1.4        2.51        pp.self.native_queued_spin_lock_slowpath
> > > > >       0.28 ±  9%      +3.8        4.06 ±  4%      +0.2        0.48            +0.4        0.68        pp.self.sctp_eat_data
> > > > >       0.00            +8.2        8.23            +0.8        0.83            +1.3        1.26        pp.self.__sk_mem_reduce_allocated
> > > > >
> > > > > And the size of 'mem_cgroup' is increased from 4224 Bytes to 4608.
> > > > 
> > > > Hi Feng, can you please try two more configurations? Take Eric's patch
> > > > of adding ____cacheline_aligned_in_smp in page_counter and for first
> > > > increase MEMCG_CHARGE_BATCH to 64 and for second increase it to 128.
> > > > Basically batch increases combined with Eric's patch.
> > > 
> > > With increasing batch to 128, the regression could be reduced to -12.4%.
> > 
> > If we're going to bump it, I wonder if we should scale it dynamically depending
> > on the size of the memory cgroup?
>  
> I think it makes sense, or also make it a configurable parameter? From 
> the test reports of 0Day, these charging/counting play critical role
> in performance (easy to see up to 60% performance effect). If user only
> wants memcg for isolating things or doesn't care charging/stats, these
> seem to be extra taxes.
> 
> For bumping to 64 or 128, universal improvement is expected with the
> only concern of accuracy.
> 
> Thanks,
> Feng
> 
> > Thanks!