Re: [PATCH v5 02/27] mm, cma: support multiple contiguous ranges, if requested

[Frank van der Linden]

On Tue, Mar 4, 2025 at 10:29 PM kernel test robot <oliver.sang@xxxxxxxxx> wrote:
>
>
>
> Hello,
>
> kernel test robot noticed a 15.1% improvement of netperf.Throughput_tps on:
>
>
> commit: a957f140831b0d42e4fdbe83cf93997ef1b51bda ("[PATCH v5 02/27] mm, cma: support multiple contiguous ranges, if requested")
> url: https://github.com/intel-lab-lkp/linux/commits/Frank-van-der-Linden/mm-cma-export-total-and-free-number-of-pages-for-CMA-areas/20250301-023339
> base: https://git.kernel.org/cgit/linux/kernel/git/torvalds/linux.git 276f98efb64a2c31c099465ace78d3054c662a0f
> patch link: https://lore.kernel.org/all/20250228182928.2645936-3-fvdl@xxxxxxxxxx/
> patch subject: [PATCH v5 02/27] mm, cma: support multiple contiguous ranges, if requested
>
> testcase: netperf
> config: x86_64-rhel-9.4
> compiler: gcc-12
> test machine: 128 threads 2 sockets Intel(R) Xeon(R) Gold 6338 CPU @ 2.00GHz (Ice Lake) with 256G memory
> parameters:
>
>         ip: ipv4
>         runtime: 300s
>         nr_threads: 200%
>         cluster: cs-localhost
>         test: TCP_CRR
>         cpufreq_governor: performance
>
>
>
>
>
>
> Details are as below:
> -------------------------------------------------------------------------------------------------->
>
>
> The kernel config and materials to reproduce are available at:
> https://download.01.org/0day-ci/archive/20250305/202503051327.e87dce82-lkp@xxxxxxxxx
>
> =========================================================================================
> cluster/compiler/cpufreq_governor/ip/kconfig/nr_threads/rootfs/runtime/tbox_group/test/testcase:
>   cs-localhost/gcc-12/performance/ipv4/x86_64-rhel-9.4/200%/debian-12-x86_64-20240206.cgz/300s/lkp-icl-2sp2/TCP_CRR/netperf
>
> commit:
>   cdc31e6532 ("mm/cma: export total and free number of pages for CMA areas")
>   a957f14083 ("mm, cma: support multiple contiguous ranges, if requested")
>
> cdc31e65328522c6 a957f140831b0d42e4fdbe83cf9
> ---------------- ---------------------------
>          %stddev     %change         %stddev
>              \          |                \
>       2.43            +0.5        2.90 ą  4%  mpstat.cpu.all.usr%
>    4718850           +15.4%    5446771        vmstat.system.cs
>      62006 ą 43%     -59.6%      25067 ą137%  numa-meminfo.node0.Mapped
>    2884295 ą 41%     -59.4%    1171696 ą135%  numa-meminfo.node0.Unevictable
>      28159 ą  2%     -17.7%      23164 ą  2%  perf-c2c.HITM.local
>       5426 ą  3%     +28.5%       6973 ą  8%  perf-c2c.HITM.remote
>      33586 ą  2%     -10.3%      30137 ą  3%  perf-c2c.HITM.total
>    5642375 ą  2%     +15.5%    6519596        sched_debug.cpu.nr_switches.avg
>    7473763 ą  4%     +18.0%    8815709 ą  2%  sched_debug.cpu.nr_switches.max
>    4352931 ą  3%     +12.7%    4906391 ą  2%  sched_debug.cpu.nr_switches.min
>    2485115 ą  6%     +31.9%    3277456 ą 11%  numa-numastat.node0.local_node
>    2526446 ą  6%     +32.8%    3356120 ą 11%  numa-numastat.node0.numa_hit
>    3522582 ą 10%     +28.7%    4535065 ą 23%  numa-numastat.node1.local_node
>    3613797 ą 10%     +27.0%    4588978 ą 22%  numa-numastat.node1.numa_hit
>      40617            +5.4%      42811 ą  5%  proc-vmstat.nr_slab_reclaimable
>    6144430 ą  4%     +29.4%    7948120 ą 16%  proc-vmstat.numa_hit
>    6011884 ą  4%     +30.0%    7815542 ą 16%  proc-vmstat.numa_local
>   26402145 ą  2%     +40.6%   37129548 ą 14%  proc-vmstat.pgalloc_normal
>   25226079           +42.1%   35834032 ą 13%  proc-vmstat.pgfree
>      15712 ą 43%     -59.6%       6348 ą137%  numa-vmstat.node0.nr_mapped
>     721073 ą 41%     -59.4%     292924 ą135%  numa-vmstat.node0.nr_unevictable
>     721073 ą 41%     -59.4%     292924 ą135%  numa-vmstat.node0.nr_zone_unevictable
>    2526848 ą  6%     +32.8%    3355902 ą 11%  numa-vmstat.node0.numa_hit
>    2485517 ą  6%     +31.9%    3277238 ą 11%  numa-vmstat.node0.numa_local
>    3614259 ą 10%     +27.0%    4589442 ą 22%  numa-vmstat.node1.numa_hit
>    3523043 ą 10%     +28.7%    4535533 ą 23%  numa-vmstat.node1.numa_local
>    1711802           +15.1%    1969470        netperf.ThroughputBoth_total_tps
>       6686           +15.1%       7693        netperf.ThroughputBoth_tps
>    1711802           +15.1%    1969470        netperf.Throughput_total_tps
>       6686           +15.1%       7693        netperf.Throughput_tps
>  4.052e+08 ą  5%     +16.7%  4.728e+08 ą  4%  netperf.time.involuntary_context_switches
>     535.88           +18.1%     633.12        netperf.time.user_time
>  3.175e+08 ą  3%     +13.9%  3.615e+08 ą  3%  netperf.time.voluntary_context_switches
>  5.135e+08           +15.1%  5.908e+08        netperf.workload
>       0.07 ą  8%     -31.3%       0.05 ą 23%  perf-sched.sch_delay.avg.ms.__cond_resched.kmem_cache_alloc_node_noprof.kmalloc_reserve.__alloc_skb.tcp_stream_alloc_skb
>       0.46 ą114%     -71.4%       0.13 ą 34%  perf-sched.sch_delay.max.ms.__cond_resched.lock_sock_nested.__inet_stream_connect.inet_stream_connect.__sys_connect
>       5.70 ą 90%   +2752.3%     162.72 ą202%  perf-sched.wait_and_delay.avg.ms.devkmsg_read.vfs_read.ksys_read.do_syscall_64
>      33.94 ą 19%     +50.3%      50.99 ą 18%  perf-sched.wait_and_delay.avg.ms.smpboot_thread_fn.kthread.ret_from_fork.ret_from_fork_asm
>      30764 ą 22%     -32.1%      20881 ą 22%  perf-sched.wait_and_delay.count.smpboot_thread_fn.kthread.ret_from_fork.ret_from_fork_asm
>       7.03 ą 60%  +11736.2%     832.16 ą150%  perf-sched.wait_and_delay.max.ms.devkmsg_read.vfs_read.ksys_read.do_syscall_64
>       0.14 ą  8%     -33.5%       0.09 ą 26%  perf-sched.wait_time.avg.ms.__cond_resched.kmem_cache_alloc_node_noprof.kmalloc_reserve.__alloc_skb.tcp_stream_alloc_skb
>       0.11 ą  8%     -14.3%       0.10 ą 11%  perf-sched.wait_time.avg.ms.__cond_resched.lock_sock_nested.inet_stream_connect.__sys_connect.__x64_sys_connect
>      33.61 ą 19%     +50.4%      50.57 ą 18%  perf-sched.wait_time.avg.ms.smpboot_thread_fn.kthread.ret_from_fork.ret_from_fork_asm
>       0.69 ą109%     -59.0%       0.28 ą 27%  perf-sched.wait_time.max.ms.__cond_resched.sk_wait_data.tcp_recvmsg_locked.tcp_recvmsg.inet_recvmsg
>       0.76           -39.9%       0.46 ą 12%  perf-stat.i.MPKI
>  3.959e+10           +14.9%   4.55e+10        perf-stat.i.branch-instructions
>       0.92            -0.0        0.90        perf-stat.i.branch-miss-rate%
>  3.564e+08           +12.7%  4.017e+08        perf-stat.i.branch-misses
>  1.561e+08           -32.2%  1.058e+08 ą 12%  perf-stat.i.cache-misses
>   6.91e+08           -33.8%  4.574e+08 ą  6%  perf-stat.i.cache-references
>    4760614           +15.5%    5496803        perf-stat.i.context-switches
>       1.54           -13.5%       1.33        perf-stat.i.cpi
>       2048           +49.1%       3054 ą  9%  perf-stat.i.cycles-between-cache-misses
>  2.084e+11           +14.9%  2.394e+11        perf-stat.i.instructions
>       0.65           +15.3%       0.75        perf-stat.i.ipc
>      37.20           +15.5%      42.97        perf-stat.i.metric.K/sec
>       0.75           -41.0%       0.44 ą 12%  perf-stat.overall.MPKI
>       0.90            -0.0        0.88        perf-stat.overall.branch-miss-rate%
>       1.54           -13.6%       1.33        perf-stat.overall.cpi
>       2060           +48.5%       3060 ą 10%  perf-stat.overall.cycles-between-cache-misses
>       0.65           +15.7%       0.75        perf-stat.overall.ipc
>  3.947e+10           +14.9%  4.536e+10        perf-stat.ps.branch-instructions
>  3.553e+08           +12.7%  4.005e+08        perf-stat.ps.branch-misses
>  1.557e+08           -32.2%  1.055e+08 ą 12%  perf-stat.ps.cache-misses
>  6.889e+08           -33.8%   4.56e+08 ą  6%  perf-stat.ps.cache-references
>    4746041           +15.5%    5479885        perf-stat.ps.context-switches
>  2.078e+11           +14.9%  2.387e+11        perf-stat.ps.instructions
>  6.363e+13           +14.9%  7.312e+13        perf-stat.total.instructions
>
>
>
>
> Disclaimer:
> Results have been estimated based on internal Intel analysis and are provided
> for informational purposes only. Any difference in system hardware or software
> design or configuration may affect actual performance.
>
>
> --
> 0-DAY CI Kernel Test Service
> https://github.com/intel/lkp-tests/wiki
>

Since there should be no functional change for existing callers of CMA
interfaces, I'm flattered by this report, but it's definitely not
these commits that cause any change in performance :-)

- Frank