some general zfs tuning (for iSCSI)

[Ronald Klop]

On Fri, 28 Jul 2017 12:56:11 +0200, Eugene M. Zheganin <emz at norma.perm.ru>  
wrote:

> Hi,
>
>
> I'm using several FreeBSD zfs installations as the iSCSI production  
> systems, they basically consist of an LSI HBA, and a JBOD with a bunch  
> of SSD disks (12-24, Intel, Toshiba or Sandisk (avoid Sandisks btw)).  
> And I observe a problem very often: gstat shows 20-30% of disk load, but  
> the system reacts very slowly: cloning a dataset takes 10 seconds,  
> similar operations aren't lightspeeding too. To my knowledge, until the  
> disks are 90-100% busy, this shouldn't happen. My systems are equipped  
> with 32-64 gigs of RAM, and the only tuning I use is limiting the ARC  
> size (in a very tender manner - at least to 16 gigs) and playing with  
> TRIM. The number of datasets is high enough - hundreds of clones, dozens  
> of snapshots, most of teh data ovjects are zvols. Pools aren't  
> overfilled, most are filled up to 60-70% (no questions about low space  
> pools, but even in this case the situation is clearer - %busy goes up in  
> the sky).
>
> So, my question is - is there some obvious zfs tuning not mentioned in  
> the Handbook ? On the other side - handbook isn't much clear on how to  
> tune zfs, it's written mostly in the manner of "these are sysctl iods  
> you can play with". Of course I have seen several ZFS tuning guides.  
> Like Opensolaris one, but they are mostly file- and  
> application-specific. Is there some special approach to tune ZFS in the  
> environment with loads of disks ? I don't know.... like tuning the vdev  
> cache or something simllar. ?
>
>
> Thanks.
>
> Eugene.


What version of FreeBSD are you running?
What is the system doing during all this?
How are your pools setup (raidz1/2/3, mirror, 3mirror)?
How is your iSCSI configured and what are the clients doing with it?
Is the data distributed evenly on all disks?
Do the clients write a lot of sync data?

I think this kind of information helps people helping you.

Regards,
Ronald.