On Fri, 4 May 2012, Peter Ross wrote:
On Thu, 3 May 2012, Tim Connors wrote:
I don't suppose anyone has ported arc_summary and zilstat to read the linux spl structures?
http://constantin.glez.de/blog/2011/02/frequently-asked-questions-about-flas...
I suspect I'd benefit greatly from l2arc in my usage (not so much zil). But I don't wanna go out and buy this yet, until I *know* it will help improve upon the 40iops I currently suffer from :)
Woot! iops were 40 or so per disk, except for the commit every 5 seconds where it went up to 500 or so (so munin's idea of average iops is mostly counting the writes). Added 16GB of usb key (a $16 usb3 key in a usb2 slot - but at least it thoroughly maxes out the usb2 bandwidth) first as secondarycache=all, and an hour ago as secondarycache=metadata (since bandwidth constrained, I just wanted to have it handle all the seeky stuff, and let the throughputty stuff go through to the grownups). And munin now tells me I've doubled the iops (writes) to the big disks because they're not spending all their time seeking anymore (got 70% hits going to my l2arc since warming up overnight)! It's consistently only using 1G of the usb key - it's allocated 15G so far, but most of that is obviously stale because most of the metadata is eventually being updated (I'm rsyncing my backuppc pool full of hardlinks from the old xfs partition, and now rsync is sitting at 20% cpu usage instead of bugger-all).
Under FreeBSD I have a lot of sysctls, see "sysctl -a | grep zfs"
Below the parts from a running machine that may interest you.
Can't you find similar sysctls or /proc/sys/whatever entries under zfsonlinux?
I have to admit I don't look at them too much at the moment because the systems work without any problems, AFAIK (and more generic nagios baed monitoring tells me).
Maybe I should setup myself a github thingy. Oh, look at that: https://github.com/spacelama/arcstat -- Tim Connors