[zfs-discuss] performance of raidz* geometries

Alex Chekholko alex at calicolabs.com
Wed Jan 24 15:54:56 EST 2018


"reading (dd and cat equivalent) of single 27 GiB file"
"        512 GiB RAM"

You'll really get some nice performance numbers with your test if you run
it several times in a row! :)

On Sun, Jan 21, 2018 at 7:27 AM, Edward Ned Harvey (zfsonlinux) via
zfs-discuss <zfs-discuss at list.zfsonlinux.org> wrote:

> > From: zfs-discuss [mailto:zfs-discuss-bounces at list.zfsonlinux.org] On
> Behalf
> > Of Markus Osterhoff via zfs-discuss
> >
> > Questions:
> > 1. With increasing # of vdevs, why does read get faster, but write get
> >    slower?
>
> I don't think your results show that. You're comparing 1.4, 1.5, 1.6...
> which are all very similar.  You're comparing 2.4, 2.5, 2.6... which are
> again, all very similar. The bigger/better question, I think is: How on
> earth is it possible that your read speed is slower than your write speed?
> That seems to defy logical explanation, unless it's the result of buffering
> writes (not an actual representation of write speed).
>
>
> > 2a Why are we so far away from XFS?
>
> For one, you're comparing a stripe with no redundancy or data integrity vs
> raidz1/raidz2. Not a very similar comparison.
>
>
> > 2b Is reading a continuous large file not the "best case" for ZFS?
>
> Depends on the pool layout. For raidz1, raidz2, it should perform pretty
> well, but it should perform even better if you have multiple threads
> reading from different parts of the file. (Or run several threads in
> parallel reading from multiple separate large files). Because a single
> thread might get stuck reading a chunk that comes from some disks,
> meanwhile some other disks are idle, and then the thread is able to request
> the next chunk, and so on.
>
>
> > 3. Our intial design was the 4*raidz2(7) layout; but it turns out that
> the
> >    available TiB (from zfs list) is 10 TiB smaller than expected; for
> >    N*raidz2(6) the free space matches our expectation.
> >    Why?
>
> Dunno. Did you do some math wrong? It's not clear from your message what
> you expected, or what you actually got. I would expect the capacity of 4x
> raidz2(7) to be 20x one disk, which should be the same for 5x raidz2(6).
> _______________________________________________
> zfs-discuss mailing list
> zfs-discuss at list.zfsonlinux.org
> http://list.zfsonlinux.org/cgi-bin/mailman/listinfo/zfs-discuss
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://list.zfsonlinux.org/pipermail/zfs-discuss/attachments/20180124/b7274e95/attachment.html>


More information about the zfs-discuss mailing list