topic Re: AdvFS Performance Question in Operating System - Tru64 Unix

AdvFS Performance Question

Mark Poeschl_2 — Wed, 23 Feb 2005 12:37:20 GMT

I've got an embedded database running on Tru64 5.1B / PK4. The storage is an EVA3000 array. Our main data volume is comprised of 5 100 GB VRAID1 virtual disks. We have bumped up the tag queue depth for HSV disks to 100. The five virtual disks make up a single AdvFS domain. Given that our load is ~ 90% read / 10% write is there anything to be gained by making five separate AdvFS domains? Any AdvFS-level data structures that would create a performance improvement from being in parallel?

Re: AdvFS Performance Question

Lisa Smith_6 — Fri, 01 Apr 2005 16:36:40 GMT

Hi Mark,

Given the current use of the domain, using 5 separate domains will not gain you any performance. However, if your load were reversed (90% write/10% read), you would gain a performance boast due to using 5 separate logs, one on each volume. The log is not used nearly as much with 90% reads. There are no other domain-wide AdvFS data structures that would give you a performance boast by using multiple domains.

Using multiple volumes, as you are, is your best bet for the best performance in your domain. This allows parallel reads across volumes.

One thing you could look into is how your "hot files" are balanced across the 5 volumes. If there are a few large files that are used all the time (your "hot files"), you want to spread these across volumes. Alternatively, you can manually spread extents of a large file across all the volumes in the domain using the "migrate" command. In effect, this is manually striping your file.

To find your hot files, use the "vfast -L hotfiles " command. You can also check I/O activity on your devices using the "iostat" and "advfsstat" commands. There are many options for these utilities that are explained in the manpages.

Lisa Smith
Tru64 AdvFS Engineering

Re: AdvFS Performance Question

Mark Poeschl_2 — Fri, 01 Apr 2005 16:58:02 GMT

Thank you Lisa!

We are pursuing striping our hot files across all the volumes, but our application vendor doesn't like the idea of de-fragging or migrating on the fly. (I know, I know) It's an embedded database called Cache from Intersystems.