ZFS performance

[MMrrTT]

Hello.

I have mirror pool. Then created dataset of 100% size and share it to PC via iSCSI.
I heard ZFS is working not very good when over 80% of space is used (96% in the latest info).
I would like to find out, if I created dataset of 100% of pool size, but dataset is 10% full, does it count and ZFS is not optimally working?
And the second question, what is the real peformance difference between <80% and >80% emptyness of the ZFS pool?

Thank you.

 

[Ericloewe]

80% is not magic, rather it's a value chosen by convention. In fact, for block storage, you should probably look more towards 50%, which already has a meaningful performance impact (Paging @jgreco and his graph). Things get progressively worse as you fill up the pool.

I would like to find out, if I created dataset of 100% of pool size, but dataset is 10% full, does it count and ZFS is not optimally working?

You mean a reservation on a dataset? No, that would not count. If you mean a zvol, it depends on what "empty" means. If the blocks were never used (or have been TRIMmed) by the iSCSI initiator, that counts as free. If the blocks were used and not TRIMmed, ZFS has no way of knowing they're free and they'll remain "used" until TRIMmed or the zvol is destroyed.

I heard ZFS is working not very good when over 80% of space is used (96% in the latest info).

If you're happy with performance, that would be fine... if you weren't at 96%. Beyond 95%, things get hairy - you should really pretend that 95% is "absolutely full, can't cram in a single extra bit". If things get to 100%, you're in for an irritating experience to recover from that.

 

[MMrrTT]

Actually, I don't have performance issues, I believe, but I think it is working not 100% perfect :)

I have auto trim on in settings.

Now I have it like this (1,79Tb free for stripe pool):

This should impact the performance?

 

[jgreco]

Paging @jgreco and his graph

A search for "Delphix" and my user name will bring up lots of appropriate discussions. Perhaps try this one.

ZFS fragmentation issues

I've said it for some time now, ZFS and iSCSI may not be a match made in heaven, due to the inevitable fragmentation caused by copy-on-write for block updates. This article shows real world results from someone who did similar work with a database, and ended up with severe fragmentation. I...

www.truenas.com

 

[Ericloewe]

I have auto trim on in settings.

Whoever is using the iSCSI volume needs to be doing the TRIMming, not TrueNAS.

This should impact the performance?

It's hard to tell from the headline numbers, as "Used" is made up of several components ("Used by Dataset", "Used by children", "Reservations", etc.), with reservations being the wrench in the machine.