Re: [LKP] [SUNRPC] 0472e47660: fsmark.app_overhead 16.0% regression

[Xing Zhengjun <zhengjun.xing@linux.intel.com>]

On 8/7/2019 3:56 PM, Xing Zhengjun wrote:
> 
> 
> On 7/24/2019 1:17 PM, Xing Zhengjun wrote:
>>
>>
>> On 7/12/2019 2:42 PM, Xing Zhengjun wrote:
>>> Hi Trond,
>>>
>>>      I attached perf-profile part big changes, hope it is useful for 
>>> analyzing the issue.
>>
>> Ping...
> 
> ping...
> 
ping...
>>
>>>
>>>
>>> In testcase: fsmark
>>> on test machine: 40 threads Intel(R) Xeon(R) CPU E5-2690 v2 @ 3.00GHz 
>>> with 384G memory
>>> with following parameters:
>>>
>>>          iterations: 20x
>>>          nr_threads: 64t
>>>          disk: 1BRD_48G
>>>          fs: xfs
>>>          fs2: nfsv4
>>>          filesize: 4M
>>>          test_size: 80G
>>>          sync_method: fsyncBeforeClose
>>>          cpufreq_governor: performance
>>>
>>> test-description: The fsmark is a file system benchmark to test 
>>> synchronous write workloads, for example, mail servers workload.
>>> test-url: https://sourceforge.net/projects/fsmark/
>>>
>>> commit:
>>>    e791f8e938 ("SUNRPC: Convert xs_send_kvec() to use iov_iter_kvec()")
>>>    0472e47660 ("SUNRPC: Convert socket page send code to use 
>>> iov_iter()")
>>>
>>> e791f8e9380d945e 0472e476604998c127f3c80d291
>>> ---------------- ---------------------------
>>>           %stddev     %change         %stddev
>>>               \          |                \
>>>      527.29           -22.6%     407.96        fsmark.files_per_sec
>>>        1.97 ± 11%      +0.9        2.88 ±  4% 
>>> perf-profile.calltrace.cycles-pp.smp_apic_timer_interrupt.apic_timer_interrupt.cpuidle_enter_state.do_idle.cpu_startup_entry 
>>>
>>>        0.00            +0.9        0.93 ±  4% 
>>> perf-profile.calltrace.cycles-pp.tcp_write_xmit.tcp_sendmsg_locked.tcp_sendmsg.sock_sendmsg.xs_sendpages 
>>>
>>>        2.11 ± 10%      +0.9        3.05 ±  4% 
>>> perf-profile.calltrace.cycles-pp.apic_timer_interrupt.cpuidle_enter_state.do_idle.cpu_startup_entry.start_secondary 
>>>
>>>        5.29 ±  2%      +1.2        6.46 ±  7% 
>>> perf-profile.calltrace.cycles-pp.svc_recv.nfsd.kthread.ret_from_fork
>>>        9.61 ±  5%      +3.1       12.70 ±  2% 
>>> perf-profile.calltrace.cycles-pp.worker_thread.kthread.ret_from_fork
>>>        9.27 ±  5%      +3.1       12.40 ±  2% 
>>> perf-profile.calltrace.cycles-pp.process_one_work.worker_thread.kthread.ret_from_fork 
>>>
>>>       34.52 ±  4%      +3.3       37.78 ±  2% 
>>> perf-profile.calltrace.cycles-pp.ret_from_fork
>>>       34.52 ±  4%      +3.3       37.78 ±  2% 
>>> perf-profile.calltrace.cycles-pp.kthread.ret_from_fork
>>>        0.00            +3.4        3.41 ±  4% 
>>> perf-profile.calltrace.cycles-pp.memcpy_erms.memcpy_from_page._copy_from_iter_full.tcp_sendmsg_locked.tcp_sendmsg 
>>>
>>>        0.00            +3.4        3.44 ±  4% 
>>> perf-profile.calltrace.cycles-pp.memcpy_from_page._copy_from_iter_full.tcp_sendmsg_locked.tcp_sendmsg.sock_sendmsg 
>>>
>>>        0.00            +3.5        3.54 ±  4% 
>>> perf-profile.calltrace.cycles-pp._copy_from_iter_full.tcp_sendmsg_locked.tcp_sendmsg.sock_sendmsg.xs_sendpages 
>>>
>>>        2.30 ±  5%      +3.7        6.02 ±  3% 
>>> perf-profile.calltrace.cycles-pp.__rpc_execute.rpc_async_schedule.process_one_work.worker_thread.kthread 
>>>
>>>        2.30 ±  5%      +3.7        6.02 ±  3% 
>>> perf-profile.calltrace.cycles-pp.rpc_async_schedule.process_one_work.worker_thread.kthread.ret_from_fork 
>>>
>>>        1.81 ±  4%      +3.8        5.59 ±  4% 
>>> perf-profile.calltrace.cycles-pp.call_transmit.__rpc_execute.rpc_async_schedule.process_one_work.worker_thread 
>>>
>>>        1.80 ±  3%      +3.8        5.59 ±  3% 
>>> perf-profile.calltrace.cycles-pp.xprt_transmit.call_transmit.__rpc_execute.rpc_async_schedule.process_one_work 
>>>
>>>        1.73 ±  4%      +3.8        5.54 ±  4% 
>>> perf-profile.calltrace.cycles-pp.xs_tcp_send_request.xprt_transmit.call_transmit.__rpc_execute.rpc_async_schedule 
>>>
>>>        1.72 ±  4%      +3.8        5.54 ±  4% 
>>> perf-profile.calltrace.cycles-pp.xs_sendpages.xs_tcp_send_request.xprt_transmit.call_transmit.__rpc_execute 
>>>
>>>        0.00            +5.4        5.42 ±  4% 
>>> perf-profile.calltrace.cycles-pp.tcp_sendmsg_locked.tcp_sendmsg.sock_sendmsg.xs_sendpages.xs_tcp_send_request 
>>>
>>>        0.00            +5.5        5.52 ±  4% 
>>> perf-profile.calltrace.cycles-pp.tcp_sendmsg.sock_sendmsg.xs_sendpages.xs_tcp_send_request.xprt_transmit 
>>>
>>>        0.00            +5.5        5.53 ±  4% 
>>> perf-profile.calltrace.cycles-pp.sock_sendmsg.xs_sendpages.xs_tcp_send_request.xprt_transmit.call_transmit 
>>>
>>>        9.61 ±  5%      +3.1       12.70 ±  2% 
>>> perf-profile.children.cycles-pp.worker_thread
>>>        9.27 ±  5%      +3.1       12.40 ±  2% 
>>> perf-profile.children.cycles-pp.process_one_work
>>>        6.19            +3.2        9.40 ±  4% 
>>> perf-profile.children.cycles-pp.memcpy_erms
>>>       34.53 ±  4%      +3.3       37.78 ±  2% 
>>> perf-profile.children.cycles-pp.ret_from_fork
>>>       34.52 ±  4%      +3.3       37.78 ±  2% 
>>> perf-profile.children.cycles-pp.kthread
>>>        0.00            +3.5        3.46 ±  4% 
>>> perf-profile.children.cycles-pp.memcpy_from_page
>>>        0.00            +3.6        3.56 ±  4% 
>>> perf-profile.children.cycles-pp._copy_from_iter_full
>>>        2.47 ±  4%      +3.7        6.18 ±  3% 
>>> perf-profile.children.cycles-pp.__rpc_execute
>>>        2.30 ±  5%      +3.7        6.02 ±  3% 
>>> perf-profile.children.cycles-pp.rpc_async_schedule
>>>        1.90 ±  4%      +3.8        5.67 ±  3% 
>>> perf-profile.children.cycles-pp.call_transmit
>>>        1.89 ±  3%      +3.8        5.66 ±  3% 
>>> perf-profile.children.cycles-pp.xprt_transmit
>>>        1.82 ±  4%      +3.8        5.62 ±  3% 
>>> perf-profile.children.cycles-pp.xs_tcp_send_request
>>>        1.81 ±  4%      +3.8        5.62 ±  3% 
>>> perf-profile.children.cycles-pp.xs_sendpages
>>>        0.21 ± 17%      +5.3        5.48 ±  4% 
>>> perf-profile.children.cycles-pp.tcp_sendmsg_locked
>>>        0.25 ± 18%      +5.3        5.59 ±  3% 
>>> perf-profile.children.cycles-pp.tcp_sendmsg
>>>        0.26 ± 16%      +5.3        5.60 ±  3% 
>>> perf-profile.children.cycles-pp.sock_sendmsg
>>>        1.19 ±  5%      +0.5        1.68 ±  3% 
>>> perf-profile.self.cycles-pp.get_page_from_freelist
>>>        6.10            +3.2        9.27 ±  4% 
>>> perf-profile.self.cycles-pp.memcpy_erms
>>>
>>>
>>> On 7/9/2019 10:39 AM, Xing Zhengjun wrote:
>>>> Hi Trond,
>>>>
>>>> On 7/8/2019 7:44 PM, Trond Myklebust wrote:
>>>>> I've asked several times now about how to interpret your results. 
>>>>> As far as I can tell from your numbers, the overhead appears to be 
>>>>> entirely contained in the NUMA section of your results.
>>>>> IOW: it would appear to be a scheduling overhead due to NUMA. I've 
>>>>> been asking whether or not that is a correct interpretation of the 
>>>>> numbers you published.
>>>> Thanks for your feedback. I used the same hardware and the same test 
>>>> parameters to test the two commits:
>>>>     e791f8e938 ("SUNRPC: Convert xs_send_kvec() to use 
>>>> iov_iter_kvec()")
>>>>     0472e47660 ("SUNRPC: Convert socket page send code to use 
>>>> iov_iter()")
>>>>
>>>> If it is caused by NUMA, why only commit 0472e47660 throughput is 
>>>> decreased? The filesystem we test is NFS, commit 0472e47660 is 
>>>> related with the network, could you help to check if have any other 
>>>> clues for the regression. Thanks.
>>>>
>>>
>>
> 

-- 
Zhengjun Xing