[opensuse] NFS perf these days?



Patrick Shanahan wrote:
* Moby<moby@xxxxxxxxxxxxxx> [05-26-12 00:04]:
The machines are primarily used as "nfs servers" and the nfs
exports are mounted for use by vmware vSphere hosts. Today, there was
massive disk corruption on all files being written to the nfs shares.
That is correct. I run an rdiff backup of a genealogy system installed on
my wife's win7 box wirelessly over nfs every night to an external usb2
disk attached to an openSUSE Tumbleweed box and see no problem/corruption
with the newer/changed files.

And I saw a post from Gregg with questions on opensuse-factory where
Tumbleweed is discussed. I suggest moving discussion there.
====
BTW == how is the read/write speed on NFS these days?

I haven't benched NFS4, but several years back I benched samba against NFS over
TCP and Samba was easily 2-3x faster. With Win7, I have benched Samba at 119MB/s
reads, and 125MB/s writes (sustained LARGE linear writes). (It easily beats out
most file transfer for max speed(was looking for a fast way to transfer large
files.. other applications may have different priorities)...

I didn't bench for small linear (or random) writes, so maybe NFS is better
in those areas, dunno...but with NFS4, I was wondering if it was any better or
if anyone had done some "objective", optimized tests of both (I tried UDP&TCP
with NFS as well as varying the read/write size to find the fastest numbers).

With Samba, on my current test setup a 16MB transfer size is optimal -- which makes
a bit of sense, considering the ethernet cards are running 9000 byte packets
(9014-header) and have a local max cache of 2000 buffers->18,000,000 -- so 16MB
would be the largest multiple of 2 I could dump into the card's buffers at one
time and allow it's TCP offload to take care of the buffer while I read another.

I.e. if I only had a 1M buffer, on the cards, or ran 1500 byte packets and the limit
was still 2k, then my optimal size might be ~2MB/buffer.

A bit annoying at the ignorance shown by various engineers -- thinking that
4k writes are still optimal -- when I asked for a increased buffer size on Tbird (and
FFox, I was told that the network only transfers a max of 1500 bytes/packet,
so anything more than that was a waste. Sigh... What can you say in the face
of such awesome intellect?! *urk*



--
To unsubscribe, e-mail: opensuse+unsubscribe@xxxxxxxxxxxx
To contact the owner, e-mail: opensuse+owner@xxxxxxxxxxxx



Relevant Pages

  • Re: (in)appropriate uses for MAXBSIZE
    ... irrelevant for nfs. ... Clustering takes a lot of CPU but doesn't affect the transfer rate much ... until the RPC read/write size was reduced. ... than the size of the buffer handed it by the buffer cache, ...
    (freebsd-arch)
  • Re: (in)appropriate uses for MAXBSIZE
    ... irrelevant for nfs. ... Clustering takes a lot of CPU but doesn't affect the transfer rate much ... nfs just does a VOP_READinto its own buffer, ... copies the results to the local file system's smaller buffers. ...
    (freebsd-arch)
  • Re: FreeBSD NFS client/Linux NFS server issue
    ... on some of our freebsd7.1 nfs clients (and it looks like we have had ... invalidating buffer cache. ...
    (freebsd-stable)
  • Re: NFS Buffering
    ... The problem is that after scp reports that it's 100% complete the ... It looks like the nfs daemon is caching write data ... sitting in a write buffer. ... buffering is done by the nfsiod. ...
    (Fedora)
  • Re: ocfs2 problem with nfs v2
    ... my guess is that this might trigger the bug. ... Today I used a Slackware 9.1 machine to mount the NFS share with NFS v2. ... buffer is too small for encoding ...
    (Linux-Kernel)