On Wed, Feb 12, 2025 at 02:00:26PM +0900, Sergey Senozhatsky wrote:
> On (25/02/07 21:09), Yosry Ahmed wrote:
> > Can we do some perf testing to make sure this custom locking is not
> > regressing performance (selfishly I'd like some zswap testing too)?
>
> So for zsmalloc I (usually) write some simple testing code which is
> triggered via sysfs (device attr) and that is completely reproducible,
> so that I compares apples to apples. In this particular case I just
> have a loop that creates objects (we don't need to compress or decompress
> anything, zsmalloc doesn't really care)
>
> - echo 1 > /sys/ ... / test_prepare
>
> for (sz = 32; sz < PAGE_SIZE; sz += 64) {
> for (i = 0; i < 4096; i++) {
> ent->handle = zs_malloc(zram->mem_pool, sz)
> list_add(ent)
> }
> }
>
>
> And now I just `perf stat` writes:
>
> - perf stat echo 1 > /sys/ ... / test_exec_old
>
> list_for_each_entry
> zs_map_object(ent->handle, ZS_MM_RO);
> zs_unmap_object(ent->handle)
>
> list_for_each_entry
> dst = zs_map_object(ent->handle, ZS_MM_WO);
> memcpy(dst, tmpbuf, ent->sz)
> zs_unmap_object(ent->handle)
>
>
>
> - perf stat echo 1 > /sys/ ... / test_exec_new
>
> list_for_each_entry
> dst = zs_obj_read_begin(ent->handle, loc);
> zs_obj_read_end(ent->handle, dst);
>
> list_for_each_entry
> zs_obj_write(ent->handle, tmpbuf, ent->sz);
>
>
> - echo 1 > /sys/ ... / test_finish
>
> free all handles and ent-s
>
>
> The nice part is that we don't depend on any of the upper layers, we
> don't even need to compress/decompress anything; we allocate objects
> of required sizes and memcpy static data there (zsmalloc doesn't have
> any opinion on that) and that's pretty much it.
>
>
> OLD API
> =======
>
> 10 runs
>
> 369,205,778 instructions # 0.80 insn per cycle
> 40,467,926 branches # 113.732 M/sec
>
> 369,002,122 instructions # 0.62 insn per cycle
> 40,426,145 branches # 189.361 M/sec
>
> 369,051,170 instructions # 0.45 insn per cycle
> 40,434,677 branches # 157.574 M/sec
>
> 369,014,522 instructions # 0.63 insn per cycle
> 40,427,754 branches # 201.464 M/sec
>
> 369,019,179 instructions # 0.64 insn per cycle
> 40,429,327 branches # 198.321 M/sec
>
> 368,973,095 instructions # 0.64 insn per cycle
> 40,419,245 branches # 234.210 M/sec
>
> 368,950,705 instructions # 0.64 insn per cycle
> 40,414,305 branches # 231.460 M/sec
>
> 369,041,288 instructions # 0.46 insn per cycle
> 40,432,599 branches # 155.576 M/sec
>
> 368,964,080 instructions # 0.67 insn per cycle
> 40,417,025 branches # 245.665 M/sec
>
> 369,036,706 instructions # 0.63 insn per cycle
> 40,430,860 branches # 204.105 M/sec
>
>
> NEW API
> =======
>
> 10 runs
>
> 265,799,293 instructions # 0.51 insn per cycle
> 29,834,567 branches # 170.281 M/sec
>
> 265,765,970 instructions # 0.55 insn per cycle
> 29,829,019 branches # 161.602 M/sec
>
> 265,764,702 instructions # 0.51 insn per cycle
> 29,828,015 branches # 189.677 M/sec
>
> 265,836,506 instructions # 0.38 insn per cycle
> 29,840,650 branches # 124.237 M/sec
>
> 265,836,061 instructions # 0.36 insn per cycle
> 29,842,285 branches # 137.670 M/sec
>
> 265,887,080 instructions # 0.37 insn per cycle
> 29,852,881 branches # 126.060 M/sec
>
> 265,769,869 instructions # 0.57 insn per cycle
> 29,829,873 branches # 210.157 M/sec
>
> 265,803,732 instructions # 0.58 insn per cycle
> 29,835,391 branches # 186.940 M/sec
>
> 265,766,624 instructions # 0.58 insn per cycle
> 29,827,537 branches # 212.609 M/sec
>
> 265,843,597 instructions # 0.57 insn per cycle
> 29,843,650 branches # 171.877 M/sec
>
>
> x old-api-insn
> + new-api-insn
> +-------------------------------------------------------------------------------------+
> |+ x|
> |+ x|
> |+ x|
> |+ x|
> |+ x|
> |+ x|
> |+ x|
> |+ x|
> |+ x|
> |+ x|
> |A A|
> +-------------------------------------------------------------------------------------+
> N Min Max Median Avg Stddev
> x 10 3.689507e+08 3.6920578e+08 3.6901918e+08 3.6902586e+08 71765.519
> + 10 2.657647e+08 2.6588708e+08 2.6580373e+08 2.6580734e+08 42187.024
> Difference at 95.0% confidence
> -1.03219e+08 +/- 55308.7
> -27.9705% +/- 0.0149878%
> (Student's t, pooled s = 58864.4)
Thanks for sharing these results, but I wonder if this will capture
regressions from locking changes (e.g. a lock being preemtible)? IIUC
this is counting the instructions executed in these paths, and that
won't change if the task gets preempted. Lock contention may be captured
as extra instructions, but I am not sure we'll directly see its effect
in terms of serialization and delays.
I think we also need some high level testing (e.g. concurrent
swapins/swapouts) to find that out. I think that's what Kairui's testing
covers.
