Random read slower with cache

[tverweij]

I just measured the disk performance without cache and with cache - every caching block size was tested.
The results:


You see that for my situation a cache block of 32k give the best results.

But what worries me is that in all cached cases the Random read 4k 32 queues, 16 threads is much slower than without cache, about 33% slower.
Is there an optimization that I overlooked, something I can do to improve this to at least the physical disk speed?

[Support]

Was your CPU usage at 100% while testing 4K@32Q16T?

The root cause is the proportion of caching processing time is relatively large in the entire 4KB data read/write process as completion time of 4KB read/write is very short. The SSD raid itself has good 4K performance, so you probably need a faster CPU and RAM to beat it.

We've been struggling for reducing cache processing time, but currently there's no better way yet.

[tverweij]

Its running on 2 CPU's with each 10 cores, each running on 2.68 Ghz without C-states, measusered on christmas day to make sure nobody was working.
So, no, CPU wasn't 100% - do you use multi threaded code for this?

But not a real big problem; in the real world tests, the cached scenario outperforms the non cached scenario without a problem

[Support]

- do you use multi threaded code for this?

Yes, we have tried and tuned with many different parameters. I hope in future we can find a better way to improve this.