Performance: ZFS v15 VS v28 on FreeNAS-8.3.0-BETA2-x64

[paleoN]

FreeNAS-8.3.0-BETA2-x64

All pools: 2 x disk mirror 4k forced

v15 pool created on FreeNAS 8.2.0-RELEASE-p1
v15 pool on FreeNAS-8.3.0-BETA2-x64
v28 pool upgraded from v15 pool on FreeNAS-8.3.0-BETA2-x64
v28 pool created on FreeNAS-8.3.0-BETA2-x64​

Using a basic, unscientific dd write from the [thread=981]performance sticky[/thread], I saw nearly the same speeds, within 5%, on 8.3.0-BETA2 with the v15 pool vs v28 pool. Those speeds were ≈ 15% slower than the v15 pool back on 8.2.0-RELEASE-p1. In other words 8.3.0-BETA2, FreeBSD 8.3 based, is ≈ 15% slower than the earlier FreeBSD 8.2 based releases. It also happens to include the much improved ZFS v28 code & saner defaults, e.g. ZFS breathing is vastly improved.

My network transfers were unaffected. My CIFs speeds still disappoint, 77Mb/sec write, 65Mb/sec read, and my FTP transfers are still near line speed Up/Down. Client-side I saw no loss in performance as the network is the limiting factor, not to mention Samba on FreeBSD.

 

[uutzinger]

Any idea where the 15% is coming from?
It might be useful to post your hard drive interface chip set/drivers.

 

[ProtoSD]

Any has idea where the 15% is coming from?
It might be useful to post your hard drive interface chip set/drivers.

It's a known issue that v28 is slower, PaleoN was just providing some verification. It's also in the release notes.

 

[uutzinger]

It's a known issue that v28 is slower, PaleoN was just providing some verification. It's also in the release notes.

Maybe I misunderstood PaleoN's text but I read the there is no measurable difference between ZFS v15 and v28 on his system but there is 15% decrease when going from v15 on FreeNAS 8.2 to v15 on FreeNAS 8.3.

 

[paleoN]

It's a known issue that v28 is slower, PaleoN was just providing some verification.

Actually, it was more of a courtesy test and I was curious myself. Also, I was tired of everyone crying about how they were never upgrading their v15 pools to v28 pools. While their concerns are valid, I felt their fears were largely overblown. Plus there is one feature I think everyone would want. I questioned whether or not they would even notice the difference. My overly simplistic testing with nearly empty mirrors seems to bear that out, YMMV.

It might be useful to post your hard drive interface chip set/drivers.

No it won't be useful as it's irrelevant. My raw disk speeds are unchanged. The raw zpool performance with a massively sequential workload is lower. However ZFS just works better as well re: saner defaults. I did a post on ZFS breathing a few months back. With those tunings on 8.2 v15 my raw zpool performance is significantly lower than what I see with the auto tuned values from 8.3 v28. None of that matters client-side.

I read the there is no measurable difference between ZFS v15 and v28 on his system

Easily, within the margin of error with the tests I did though the v28 pool finished last all the times.

but there is 15% decrease when going from v15 on FreeNAS 8.2 to v15 on FreeNAS 8.3.

Yes, raw massively sequential zpool performance, i.e. not a real world workload, fell with a v15 pool but different FreeNAS versions.

By default 8.3 syncs txg commits more frequently and doesn't allow them to grow as large. This along with the other innumerable ZFS bug fixes and ZFS code changes causes the 'reduced' performance. Most systems have sufficient disk bandwidth. It's the latencies that you feel. 8.3 seems to do much better with the latencies.

My recommendation is to throw out any tuneables you have, set vfs.zfs.arc_max, turn on Prefetch, there were some improvements, and reevaluate.

 

[uutzinger]

Yes, raw massively sequential zpool performance, i.e. not a real world workload, fell with a v15 pool but different FreeNAS versions.

I use my system for 70 Gigabyte single file transfer from host A to host B (research data). I assume that is massively sequential. Do your tuning recommendations hold up under such conditions?

Thanks for previous reply.

 

[paleoN]

I use my system for 70 Gigabyte single file transfer from host A to host B (research data). I assume that is massively sequential.

Is host A the only writer? Is host A only writing and not reading? If so:

My network transfers were unaffected.

Unless you have different networking than I suspect.

Do your tuning recommendations hold up under such conditions?

What recommendations? The one to delete any & all tuneables and start over tuning on 8.3? Yes.