On Wed, Dec 29, 2021 at 2:34 PM Dirk Müller <dmueller@xxxxxxx> wrote:
>
> In commit fe5cbc6e06c7 ("md/raid6 algorithms: delta syndrome functions")
> a xor_syndrome() has been added for optimized syndrome calculation and
> being added also to to the raid6_choose_gen() function. However, the
> result of the xor_syndrome() benchmarking was intentionally discarded
> and did not influence the choice.
>
> We can optimize raid6_choose_gen() without changing its behavior by
> only benchmarking the chosen xor_syndrome() variant and printing it
> output, rather than benchmarking also the ones that are discarded and
> never influence the outcome.
>
> For x86_64, this removes 8 out of 18 benchmark loops which each take
> 16 jiffies, so up to 160 jiffies saved on module load (640ms on a 250HZ
> kernel).
How critical is 160 (or 128?) jiffies saving here? I think some users use these
information from dmesg, which shows an overview of all these algorithms:
[ 21.629694] raid6: avx2x4 gen() 18081 MB/s
[ 21.799692] raid6: avx2x4 xor() 8568 MB/s
[ 21.969692] raid6: avx2x2 gen() 15060 MB/s
[ 22.139792] raid6: avx2x2 xor() 9665 MB/s
[ 22.309718] raid6: avx2x1 gen() 9223 MB/s
[ 22.479691] raid6: avx2x1 xor() 6809 MB/s
[ 22.649691] raid6: sse2x4 gen() 9647 MB/s
[ 22.819691] raid6: sse2x4 xor() 6229 MB/s
[ 22.989690] raid6: sse2x2 gen() 7276 MB/s
[ 23.159690] raid6: sse2x2 xor() 5400 MB/s
[ 23.329761] raid6: sse2x1 gen() 4734 MB/s
[ 23.499693] raid6: sse2x1 xor() 3565 MB/s
[ 23.500225] raid6: using algorithm avx2x4 gen() 18081 MB/s
[ 23.500886] raid6: .... xor() 8568 MB/s, rmw enabled
If it is critical for some use cases, maybe we can gate the change with a
CONFIG?
Thanks,
Song
[...]
