Hi Dmitry, On 2022/10/11 23:58, Dmitry Kozlyuk wrote: > 2022-10-11 12:17 (UTC+0000), Chengwen Feng: >> This patch adds memarea to malloc_perf_autotest. >> >> Test platform: Kunpeng920 >> Test command: dpdk-test -a 0000:7d:00.3 -l 10-12 >> Test result: >> USER1: Performance: rte_memarea >> USER1: Size (B) Runs Alloc (us) Free (us) Total (us) memset (us) >> USER1: 64 10000 0.03 0.03 0.06 0.01 >> USER1: 128 10000 0.02 0.03 0.05 0.01 >> USER1: 1024 10000 0.03 0.05 0.07 0.20 >> USER1: 4096 10000 0.03 0.05 0.07 0.34 >> USER1: 65536 10000 0.10 0.08 0.18 2.14 >> USER1: 1048576 644 0.10 0.04 0.14 29.07 >> USER1: 2097152 322 0.10 0.04 0.14 57.50 >> USER1: 4194304 161 0.12 0.04 0.15 114.50 >> USER1: 16777216 40 0.11 0.04 0.15 456.09 >> USER1: 1073741824 Interrupted: out of memory. [1] >> >> Compared with rte_malloc: >> USER1: Performance: rte_malloc >> USER1: Size (B) Runs Alloc (us) Free (us) Total (us) memset (us) >> USER1: 64 10000 0.14 0.07 0.21 0.01 >> USER1: 128 10000 0.10 0.05 0.15 0.01 >> USER1: 1024 10000 0.11 0.18 0.29 0.21 >> USER1: 4096 10000 0.13 0.39 0.53 0.35 >> USER1: 65536 10000 0.17 2.27 2.44 2.15 >> USER1: 1048576 10000 37.21 71.63 108.84 29.08 >> USER1: 2097152 10000 8831.15 160.02 8991.17 63.52 >> USER1: 4194304 10000 47131.88 413.75 47545.62 173.79 >> USER1: 16777216 4221 119604.60 2209.73 121814.34 964.42 >> USER1: 1073741824 31 335058.32 223369.31 558427.63 62440.87 >> >> [1] The total-size of the memarea is restricted to avoid creation >> failed. > > This is not a fair comparison: > rte_malloc time includes obtaining memory from the system.
Yes, but I want to keep this patch, at least we know the different. > I think that memarea should have a dedicated benchmark, > because eventually it will be interesting to compare memarea > with different sources and algorithms. It may take a long time to reach a benchmark that everyone agrees with. I will try after this patch set upstreamed. > It will be also possible to add DPDK allocator to the comparison > by running it for an isolated heap that doesn't grow. > (In some distant future it would be cool to make DPDK allocator pluggable!) > Some shared code between this benchmark and the new one can be factored out. > . >
