sock: introduce dynamic zerocopy according to data size
MSG_ZEROCOPY is not always effective as mentioned in
https://www.kernel.org/doc/html/v4.15/networking/msg_zerocopy.html.
Currently in spdk, once we enable sendmsg zerocopy, then all data
transferred through _sock_flush are sent with zerocopy, and vice
versa. Here dynamic zerocopy is introduced to allow data sent with
MSG_ZEROCOPY or not according to its size, which can be enabled by
setting "enable_dynamic_zerocopy" as true.
Test with 16 P4610 NVMe SSD, 2 initiators, target's and initiators'
configurations are the same as spdk report:
https://ci.spdk.io/download/performance-reports/SPDK_tcp_perf_report_2104.pdf
For posix socket, rw_percent=0(randwrite), it has 1.9%~8.3% performance boost
tested with target 1~40 cpu cores and qdepth=128,256,512. And it has no obvious
influence when read percentage is greater than 50%.
For uring socket, rw_percent=0(randwrite), it has 1.8%~7.9% performance boost
tested with target 1~40 cpu cores and qdepth=128,256,512. And it still has
1%~7% improvement when read percentage is greater than 50%.
The following is part of the detailed data.
posix:
qdepth=128
rw_percent 0 | 30
cpu origin thisPatch opt | origin thisPatch opt
1 286.5 298.5 4.19% 307 304.15 -0.93%
4 1042.5 1107 6.19% 1135.5 1136 0.04%
8 1952.5 2058 5.40% 2170.5 2170.5 0.00%
12 2658.5 2879 8.29% 3042 3046 0.13%
16 3247.5 3460.5 6.56% 3793.5 3775 -0.49%
24 4232.5 4459.5 5.36% 4614.5 4756.5 3.08%
32 4810 5095 5.93% 4488 4845 7.95%
40 5306.5 5435 2.42% 4427.5 4902 10.72%
qdepth=512
rw_percent 0 | 30
cpu origin thisPatch opt | origin thisPatch opt
1 275 287 4.36% 294.4 295.45 0.36%
4 979 1041 6.33% 1073 1083.5 0.98%
8 1822.5 1914.5 5.05% 2030.5 2018.5 -0.59%
12 2441 2598.5 6.45% 2808.5 2779.5 -1.03%
16 2920.5 3109.5 6.47% 3455 3411.5 -1.26%
24 3709 3972.5 7.10% 4483.5 4502.5 0.42%
32 4225.5 4532.5 7.27% 4463.5 4733 6.04%
40 4790.5 4884.5 1.96% 4427 4904.5 10.79%
uring:
qdepth=128
rw_percent 0 | 30
cpu origin thisPatch opt | origin thisPatch opt
1 270.5 287.5 6.28% 295.75 304.75 3.04%
4 1018.5 1089.5 6.97% 1119.5 1156.5 3.31%
8 1907 2055 7.76% 2127 2211.5 3.97%
12 2614 2801 7.15% 2982.5 3061.5 2.65%
16 3169.5 3420 7.90% 3654.5 3781.5 3.48%
24 4109.5 4414 7.41% 4691.5 4750.5 1.26%
32 4752.5 4908 3.27% 4494 4825.5 7.38%
40 5233.5 5327 1.79% 4374.5 4891 11.81%
qdepth=512
rw_percent 0 | 30
cpu origin thisPatch opt | origin thisPatch opt
1 259.95 276 6.17% 286.65 294.8 2.84%
4 955 1021 6.91% 1070.5 1100 2.76%
8 1772 1903.5 7.42% 1992.5 2077.5 4.27%
12 2380.5 2543.5 6.85% 2752.5 2860 3.91%
16 2920.5 3099 6.11% 3391.5 3540 4.38%
24 3697 3912 5.82% 4401 4637 5.36%
32 4256.5 4454.5 4.65% 4516 4777 5.78%
40 4707 4968.5 5.56% 4400.5 4933 12.10%
Signed-off-by: Richael Zhuang <richael.zhuang@arm.com>
Change-Id: I730dcf89ed2bf3efe91586421a89045fc11c81f0
Reviewed-on: https://review.spdk.io/gerrit/c/spdk/spdk/+/12210
Community-CI: Broadcom CI <spdk-ci.pdl@broadcom.com>
Community-CI: Mellanox Build Bot
Tested-by: SPDK CI Jenkins <sys_sgci@intel.com>
Reviewed-by: Aleksey Marchuk <alexeymar@nvidia.com>
Reviewed-by: Ben Walker <benjamin.walker@intel.com>
9bff828f99