Low nfs write throughput

[John Baldwin]

On Monday, November 28, 2011 7:12:39 pm Daryl Sayers wrote:
> >>>>> "Bengt" == Bengt Ahlgren <bengta at sics.se> writes:
> 
> > Daryl Sayers <daryl at ci.com.au> writes:
> >> Can anyone suggest why I am getting poor write performance from my nfs setup.
> >> I have 2 x FreeBSD 8.2-STABLE i386 machines with ASUS P5B-plus mother boards,
> >> 4G mem and Dual core 3g processor using 147G 15k Seagate SAS drives with
> >> onboard Gb network cards connected to an idle network. The results below show
> >> that I get nearly 100Mb/s with a dd over rsh but only 15Mbs using nfs. It
> >> improves if I use async but a smbfs mount still beats it. I am using the same
> >> file, source and destinations for all tests. I have tried alternate Network
> >> cards with no resulting benefit.
> 
> > [...]
> 
> >> Looking at a systat -v on the destination I see that the nfs test does not
> >> exceed 16KB/t with 100% busy where the other tests reach up to 128KB/t.
> >> For the record I get reads of 22Mb/s without and 77Mb/s with async turned on
> >> for the nfs mount.
> 
> > On an UFS filesystem you get NFS writes with the same size as the
> > filesystem blocksize.  So an easy way to improve performance is to
> > create a filesystem with larger blocks.  I accidentally found this out
> > when I had two NFS exported filesystems from the same box with 16K and
> > 64K blocksizes respectively.
> 
> > (Larger blocksize also tremendously improves the performance of UFS
> > snapshots!)
> 
> Thanks to all that answered. I did try the 'sysctl -w vfs.nfsrv.async=1' with
> no reportable change in performance. We are using a UFS2 filesystem so the
> zfs command was not required. I did not try the patch as we would like to stay
> as standard as possible but will upgrade if the patch is released in new
> kernel.

If you can test the patch then it is something I will likely put into the
next release.  I have already tested it as far as robustness locally, what
I don't have are good performance tests.  It would really be helpful if you
were able to test it.

> Thanks Bengt for the suggestion of block size. Increasing the block size to
> 64k made a significant improvement to performance.

In theory the patch might have given you similar gains.  During my simple tests
I was able to raise the average I/O size in iostat to 70 to 80k from 16k.

-- 
John Baldwin