Some follow-up stuff:
I ran the sequential write tests using dd as in:
# time dd if=/dev/zero of=/opt/test/test.img bs=1M count=3000
with three settings that aren't XFS-related but are NFS-specific. Namely,
the wsize and rsize settings of NFSv3.
- 8192 - - 32768 - - 65535 -
real 8m23.713s 8m0.675s 8m39.662s
user 0m0.010s 0m0.020s 0m0.040s
sys 0m34.890s 0m35.350s 0m35.480s
The differences don't seem to be that significant, but I'm going with
32768 since my system is using pure NFSv3 and not NFSv2. Besides, this is
set in the client side so an NFSv2 client can always use 8192 for both
rsize and wsize. Anyway, this is off-topic for the XFS list (sorry).
An observation about disk activity:
Disk activity is pretty continuous, which is good. However there are
situations where disk activity will stop but network activity continuous.
I presume this is filling up some buffer. Then disk activity will pick up
and network activity will stop for awhile. I presume aggressive flushing
of buffers is going on. This doesn't look quite like "streaming" to me.
I'm not complaining, but just thought I'd relay the tidbit in the hopes of
getting some enlightenment.
Some notes on my system:
o I'm using RAID5, and even if this is hardware RAID, RAID5 writes are
not so good. I should have gone RAID10, but it's too late to shift now.
o I created the filesystems using "-l size=32768b" in mkfs.xfs
o I mount the filesystems using
"rw,noatime,nodiratime,logbufs=8,biosize=16,osyncisdsync"
:-)
"Very happy with NFS+XFS"
--> Jijo
--
Federico Sevilla III :: jijo@xxxxxxxxxxxxxxxxxxxx
Network Administrator :: The Leather Collection, Inc.
GnuPG Key: <http://jijo.leathercollection.ph/jijo.gpg>
|