Comments on: How multiple disks can benefit for single client workload ?

By: David Holoboff

David Holoboff — Tue, 29 Apr 2008 14:34:18 +0000

Peter,

Do you have any posts that compare filesystem and IO schedulers, and perhaps a SATA vs. SCSI comparison?

By: Marki

Marki — Tue, 29 Apr 2008 08:37:36 +0000

And what about SAN disk arrays like EVA? They use all available disk spindles for every LUN – so when you have 30x 73 GB 15k disks and create 100 GB LUN, it will distributed over all 30 disks. I don’t remember the chunk size now (it is 32 MB, but I think the piece which is stripped is 2 MB). So there is very big chance that every request will come to different disk. Also you have mirroring, so every chunk can be read from 2 drives.

By: peter

peter — Sun, 27 Apr 2008 16:40:11 +0000

Matt sure,

There are many different issues people run into and I wanted to write about particular one, though single thread is important one because it is something which often comes as surprise. Haven not you ever seen Slaves which are unable to keepup with master even if you do not have any side load on them or people trying to speedup long running queries but getting expensive hardware and failing to see substantial benefits ?

In many cases you would of course see multiple clients but it is performance of single thread is causing you a problem – sometimes this would be because of competition with other threads but often it is simply limitations of how well single thread can use multiple disks.

By: Matt yonkovit

Matt yonkovit — Sun, 27 Apr 2008 14:03:04 +0000

Nice follow-up peter. For me I am looking beyond the single thread IO bottleneck, and mainly concerned with customer workloads. In these cases I can see AWAIT times that approach some insane levels. In fact I want to say maybe 80% of the people who I deal with have substandard latency coming from their IO subsystem. I don’t think I ever have encountered a single threaded latency issue. So mostly I see High queues, tons of requests, high wait times, etc. Sometimes adding more memory gets the job done, but many times they are maxed out.

So its all about the IO ( contention ) !