First thing I would check is that the GPFS block size is a multiple of a
full RAID stripe. It’s been a while since I worked with SVC/FlashSystem
performance, but this has been my main issue. So, 8+2p with the default
128KB «chunk size» would work with 1 MB or larger block size.

The other thing was that it’s important to disable prefetching (chsystem
-cacheprefetch off), as it will always be prefetching the wrong data
because of how GPFS scatters the blocks.

And.. on linux side there’s some max device transfersize setting that has
had huge impact on some systems.. But the exact setting escapes me right
now..


HTH


  -jf


tir. 23. jan. 2024 kl. 15:05 skrev Petr Plodík <[email protected]>:

> Hi,
>
> we have GPFS cluster with two IBM FlashSystem 7300 systems with HD
> expansion and 80x 12TB HDD each (in DRAID 8+P+Q), 3 GPFS servers connected
> via 32G FC. We are doing performance tuning on sequential writes to HDDs
> and seeing suboptimal performance. After several tests, it turns out, that
> the bottleneck seems to be the single HDD write performance, which is below
> 40MB/s and one would expect at least 100MB/s.
>
> Does anyone have experiences with IBM flashsystem sequential write
> performance tuning or has these arrays in the infrastructure? We would
> really appreciate any help/explanation.
>
> Thank you!
>
> Petr Plodik
> M Computers s.r.o.
> [email protected]
>
>
>
> _______________________________________________
> gpfsug-discuss mailing list
> gpfsug-discuss at gpfsug.org
> http://gpfsug.org/mailman/listinfo/gpfsug-discuss_gpfsug.org
>
_______________________________________________
gpfsug-discuss mailing list
gpfsug-discuss at gpfsug.org
http://gpfsug.org/mailman/listinfo/gpfsug-discuss_gpfsug.org

Reply via email to