Ross - ZFS IO stats per dataset

New About Yours API Help
1.4 KB, Plain text
You asked for questions for the pre-recorded episodes coming up, so here you go:

I am trying to optimize performance for MariaDB, MongoDB, and a Tomcat app mainly writing to a single pool. I have read all the articles about getting the right recordsize, etc., but after a previous round of tuning I actually think performance may have gone slightly down. The switch was from 8k to 16k blocks on the InnoDB dataset and creating a new dataset for the InnoDB logs with 128k blocks. Both datasets are lz4 compressed. Now, it is possible that the DB load increased slightly since the change, but my 24h average system-wide iowait CPU % went from 2-3% to 3-5%, so nearly double. (Yes, this is ZFS-on-Linux.)

Is there any way, using the ZFS tools to monitor either iops per dataset, like ‘zpool stat’? Is there a ZFS way to see application write request buffer sizes? I understand the latter is probably OS-dependent, with dtrace being the way on FreeBSD, but I’m wondering if there is a “ZFS-native” way.

In your answer, if there are any BSD-specific ZFS features that are relevant, please do enlighten me, so I know what’s missing from ZoL! If dtrace is the answer, I guess I’ll just have to wait a little longer. :-)

Thanks for the show and all the learning opportunities it provides. I enjoyed meeting you both recently at AsiaBSDCon. I’m working on being able to also attend BSDCan, so we’ll see.

Pasted 3 weeks, 4 days ago — Expires in 340 days