On Tue, Oct 22, 2019 at 4:24 PM Eric Dumazet <eduma...@google.com> wrote: > > On Tue, Oct 22, 2019 at 4:11 PM Cong Wang <xiyou.wangc...@gmail.com> wrote: > > > > Currently RTO, TLP and PROBE0 all share a same timer instance > > in kernel and use icsk->icsk_pending to dispatch the work. > > This causes spinlock contention when resetting the timer is > > too frequent, as clearly shown in the perf report: > > > > 61.72% 61.71% swapper [kernel.kallsyms] > > [k] queued_spin_lock_slowpath > > ... > > - 58.83% tcp_v4_rcv > > - 58.80% tcp_v4_do_rcv > > - 58.80% tcp_rcv_established > > - 52.88% __tcp_push_pending_frames > > - 52.88% tcp_write_xmit > > - 28.16% tcp_event_new_data_sent > > - 28.15% sk_reset_timer > > + mod_timer > > - 24.68% tcp_schedule_loss_probe > > - 24.68% sk_reset_timer > > + 24.68% mod_timer > > > > This patch decouples TLP timer from RTO timer by adding a new > > timer instance but still uses icsk->icsk_pending to dispatch, > > in order to minimize the risk of this patch. > > > > After this patch, the CPU time spent in tcp_write_xmit() reduced > > down to 10.92%. > > What is the exact benchmark you are running ? > > We never saw any contention like that, so lets make sure you are not > working around another issue.
I simply ran 256 parallel netperf with 128 CPU's to trigger this spinlock contention, 100% reproducible here. A single netperf TCP_RR could _also_ confirm the improvement: Before patch: $ netperf -H XXX -t TCP_RR -l 20 MIGRATED TCP REQUEST/RESPONSE TEST from 0.0.0.0 (0.0.0.0) port 0 AF_INET to XXX () port 0 AF_INET : first burst 0 Local /Remote Socket Size Request Resp. Elapsed Trans. Send Recv Size Size Time Rate bytes Bytes bytes bytes secs. per sec 655360 873800 1 1 20.00 17665.59 655360 873800 After patch: $ netperf -H XXX -t TCP_RR -l 20 MIGRATED TCP REQUEST/RESPONSE TEST from 0.0.0.0 (0.0.0.0) port 0 AF_INET to XXX () port 0 AF_INET : first burst 0 Local /Remote Socket Size Request Resp. Elapsed Trans. Send Recv Size Size Time Rate bytes Bytes bytes bytes secs. per sec 655360 873800 1 1 20.00 18829.31 655360 873800 (I have run it for multiple times, just pick a median one here.) The difference can also be observed by turning off/on TLP without patch. Thanks.
