Terrible NFS4 performance: FreeBSD 9.1 + UFS/ZFS + AWS EC2
Julian Elischer
julian at freebsd.org
Wed Jul 31 07:22:40 UTC 2013
On 7/31/13 9:08 AM, Berend de Boer wrote:
>>>>>> "Rick" == Rick Macklem <rmacklem at uoguelph.ca> writes:
> Rick> I think you mentioned that you were using a Linux client,
> Rick> but not what version. I'd suggest a recent kernel from
> Rick> kernel.org. (Fedora tracks updates/fixes for NFSv4 pretty
> Rick> closely, so the newest Fedora release should be pretty
> Rick> current.)
>
> This was Ubuntu 10.04 LTS.
>
> Have just tried a FreeBSD 9.1 client. Similar numbers. NFSv3 is about
> 30% slower on FreeBSD than Linux: 3m30s versus 2m10s. NFSv4 has the
> same terribly slow performance, i.e. 21m56s for the same test.
>
> Interestingly, the nfsd cpu usage doesn't rise as high as with
> Linux. But goes up to 20% (instead of over 50%).
>
> I had a look at collectd measurements as well, one cpu on the FreeBSD
> server is spending a lot of time in IRQ (whatever that means).
>
> BTS, this was a FreeBSD NFS4 out-of-the-box server, not with the patch
> (as the patch didn't do that much for me, it did some, but performance
> was still 8 times slower than nfs3).
>
>
> Rick> All I can suggest is capturing packets and then emailing be
> Rick> the captured packet trace. You can use tcpdump to do the
> Rick> capture, since wireshark will understand it: # tcpdump -s 0
> Rick> -w <file>.pcap host <client-host> and then emailing me
> Rick> <file>.pcap.
>
> Rick> I can take a look at the packet capture and maybe see what
> Rick> is going on.
>
> Will email them shortly.
Recent evidence with AWS is suggesting that the NOADAPTIVE_XXX options
in the
XENHVM kernel are now seriously hindering AWS performance all over the
place.
make sure you have tried with these options removed.
> --
> All the best,
>
> Berend de Boer
>
>
> ------------------------------------------------------
> Awesome Drupal hosting: https://www.xplainhosting.com/
More information about the freebsd-fs
mailing list