Robin Humble <rjh@xxxxxxxxxxxxxxxxxxxxxxxxxxx> wrote:
> However, (and I guess you knew this was coming :-) I tried writing to
> an XFS partition over NFS again - this time using work machines
> instead of home, and the performance was a still a bit of a worry
> to say the least.
> machine/config this time:
> dual celeron 500, ASUS bp6, 512M ram, 4x maxtor 60G/7200rpm drives,
> 2x promise ultra100 controller cards (each drive on its own controller),
> RH7.1beta, 2.4.2-XFSsmp-3mar2001 kernel. The network is 100Mbit
> ethernet (3c905 I think). Reiserfs was not timed as it was unable to
> make >2G files(!!)
> time dd if=/dev/zero of=bigFile bs=1024k count=2500
> writes to local disk writes over nfs notes
> xfs 16 MB/s 0.7 MB/s load ~90-100% per cpu + (*)
> xfs over RAID0 37 MB/s 1 MB/s ""
> xfs over RAID5 didn't try yet
> ext2 17 MB/s 8 MB/s load ~30-40% per cpu
> ext2 over RAID0 38 MB/s 8 MB/s ""
> ext2 over RAID5 14 MB/s 6 MB/s "" + (**)
i just did a very simple test on the setup i can us right now from
remote and end up at about 2mb/sec (server: up 333mhz pII, 128mb,
6gb udma2 disk - so - no big hardware, ah - i forgot: k6-2/400, 64mb
solaris8 ix86 client) - looks like something is definitely not ok
with your setup (otherwise i would not get better results with my
much slower hardware) - so - some things to try:
* did you try kiocluster's (only work with ide or scsi not with md
or lvm) - no idea if it might help -> mount -o kiocluster ...
* the other good lines posted often here (but i don't think they
will help here) -> mount -o logbufs=4,logbsize=32768 ...
* creating the filesystem with a bigger journal is also a good idea
for real workload (ok - i think not in this case too - because
here is not much metadata to journal :-)
-> mkfs -t xfs -l size=32768b ...
* did you try the same with an up kernel? - maybe some smp locking
the linux architects
tel: +49-30-308806-13 fax: -77 http://www.innominate.com