Re: NFS client (10x) performance regression 2.6.14.7 -> 2.6.15

From: Bill Rugolsky Jr.
Date: Fri Mar 31 2006 - 11:32:43 EST


On Fri, Mar 31, 2006 at 05:24:01PM +0200, Jakob Oestergaard wrote:
> With a local export/mount I see very few reads running the test :/
>
> Could this be a timing issue? That pages in the cache are invalidated
> after maybe less than 100 microseconds which is roughly the round-trip
> time on the gigabit network connecting my server and client?

I don't see any problem over loopback either. Perhaps I'll try increasing
delays with tc netem. That will have to wait until later.

In any case, I tried on our NetApp FAS-250, and I see the same problem:
Mount options are:
netapp:/home /nfs/netapp/home nfs rw,v3,rsize=32768,wsize=32768,hard,intr,tcp,lock,addr=netapp 0 0

I don't have stat zeroing, so I rolled some scripts. Sorry, these are
not vanilla kernels; if you don't nail this today, I'll try and find time
for git bisection over the weekend.

-Bill

==

My production FC1 2.6.10 kernel:

ti64: rm -f testfile ; uname -a ; ( /tmp/nfsstat.sh ; time /tmp/nfsbench ; /tmp/nfsstat.sh ) | /tmp/nfsstatdiff.awk
Linux ti64 2.6.10-0.ti.4.fc1smp #1 SMP Wed Jan 19 12:15:34 EST 2005 i686 athlon i386 GNU/Linux

real 0m0.215s
user 0m0.000s
sys 0m0.093s
null: 0 getattr: 2 setattr: 0 lookup: 1 access: 0 readlink: 0 read: 1 write: 321 create: 1 mkdir: 0 symlink: 0 mknod: 0 remove: 0 rmdir: 0 rename: 0 link: 0 readdir: 0 readdirplus: 0 fsstat: 0 fsinfo: 0 pathconf: 0 commit: 0

===

DaveJ's FC4 2.6.15.3-based errata kernel, mounted:

ti63: rm -f testfile ; uname -a ; ( /tmp/nfsstat.sh ; time /tmp/nfsbench ; /tmp/nfsstat.sh ) | /tmp/nfsstatdiff.awk
Linux ti63 2.6.15-1.1831_FC4smp #1 SMP Tue Feb 7 13:48:31 EST 2006 i686 athlon i386 GNU/Linux

real 0m3.242s
user 0m0.000s
sys 0m0.360s
null: 0 getattr: 2561 setattr: 0 lookup: 1 access: 1 readlink: 0 read: 7040 write: 3200 create: 1 mkdir: 0 symlink: 0 mknod: 0 remove: 0 rmdir: 0 rename: 0 link: 0 readdir: 0 readdirplus: 0 fsstat: 0 fsinfo: 0 pathconf: 0 commit: 0


-Bill
-
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/