Description of Problem:
(from my post to comp.os.linux.redhat):
The short version: why has NFS client support gotten so much worse
between RH 7.2 and 7.3, and what can we do about it?
Alright, some background: we're a mixed-platform Unix shop, with
Solaris Sparc servers and a mix of primarily Linux and Solaris desktops
(with a few of just about everything else for good measure - IRIX, Tru64,
HP-UX, AIX, OS X, Win2k). The network is NFS/NIS using the automounter,
and has worked quite well historically.
While most of our Linux boxes are running RedHat 7.2, I've begun
installing a few limited RH 7.3 test boxes in the group for testing
purposes. So far, it's going great except for one thing: the NFS
performance. If the systems attempt to copy any large files or perform
any non-trivial NFS task, they get errors of the sort of:
"nfs: task xxxx can't get a request slot"
...and the task doesn't really complete. Even when it does work,
performance is terrible - copying a 256M file to local disk takes minutes
instead of seconds on similar machines (10:35.43 on a RH 7.3 box, compared
to 0:50.18 on 7.2).
I have tried patching the 7.3 boxes to the most recent kernel
RPM (2.4.18-5), which hasn't changed anything significantly. On searching
Usenet, I've found word that I should be lowering the default wsize and
rsize to 8k or 16k; I don't consider this an acceptable alternative, given
that we use the automounter and any changes I make there will affect all
other systems on our network.
Otherwise, I'm stumped. And I'm really tempted to give RH 7.3 a
miss, which seems a shame but apparently necessary.
Version-Release number of selected component (if applicable):
7.3, kernels 2.4.18-3 and 2.4.18-5 (either)
Steps to Reproduce:
On two (similar but not identical) systems:
# mount reno:/export/reno/ext36 /mnt
# time cp /mnt/isoimages/valhalla-i386-disc1.iso /scratch
RH 7.2 (2.4.9-34) results:
0.060u 16.580s 2:15.84 12.2% 0+0k 0+0io 103pf+0w
RH 7.3 (2.4.18-5) results:
0.070u 62.730s 5:47.73 18.0% 0+0k 0+0io 100pf+0w
We believe this problem is fixed in the 7.3 errata kernel which should be coming
out within a week or so.