ZFS performance bottlenecks: CPU or RAM or anything else?

[Steven Hartland]

Not to mention it's so easy to cripple performance with a bad bios setup
this could easily be a simple setup issue.

I had an issue the other day where a 4ghz Intel CPU couldn't process video
transcode in real time which turned out to be a power saving option in the
bios that was utterly destroying performance by running the CPU at 800Mhz
instead of 4Ghz. Everything else seemed fine with nothing was using more
then a few present of CPU. Disabling power saving fixed the issue. This
issue was not present on a much lower power / older box simply because it
didn't have the advanced power saving options.

I'm not saying this was the case in these tests but simply providing a
comcrete example that it's sometimes hard to get like for like comparisons
even for what should be simple tests.

On Tuesday, 17 May 2016, Matthew D. Fuller <fullermd at over-yonder.net> wrote:

> On Tue, May 17, 2016 at 02:16:16PM -0700 I heard the voice of
> Freddie Cash, and lo! it spake thus:
> >
> > They're not asking for ways to improve the performance of a
> > raidz-based pool; they're asking why they get different performance
> > metrics from the exact same pool when they change the CPU and RAM.
>
> More specifically, as I read it, different performance in a very
> specific metric; single-thread linear bulk writes.  That doesn't seem
> like it would benefit heavily from a lot of cores available, or from
> RAM bandwidth or size above a pretty low threshold.
>
> Of course, it's not just changing the CPU and RAM; it's also the
> motherboard, and possibly the HBA (at least the bus the HBA is on, if
> it's a card being transplanted with the pool).  And the Core 2 would
> be back in the plain-old FSB era, so RAM access would be competing
> with the disk IO on the bus.
>
>
> --
> Matthew Fuller     (MF4839)   |  fullermd at over-yonder.net <javascript:;>
> Systems/Network Administrator |  http://www.over-yonder.net/~fullermd/
>            On the Internet, nobody can hear you scream.
>