I'm a bit stymied over some iostat comparisons I've been making
between a couple different kernels. If someone could help me with the
numbers...
I've got a Linux server fronting a RAID array (I mentioned this a week
or so ago, but the implementation details are probably unimportant).
I did some testing with the SGI-built XFS-enabled version of Red Hat's
2.4.18 kernel (2.4.18-18SGI_XFS_1.2pre5smp).
I pointed a half-dozen NFS clients at the exported filesystem and had
them run iozone simultaneously. Here's a representative except of the
server's iostat output while the test was going on (sorry for the line
wrapping):
Device: rrqm/s wrqm/s r/s w/s rsec/s wsec/s rkB/s wkB/s
avgrq-sz avgqu-sz await svctm %util
/dev/sda 0.00 4040.53 0.00 311.57 0.00 32864.67 0.00
16432.33 105.48 194.33 21.00 3.21 100.00
Device: rrqm/s wrqm/s r/s w/s rsec/s wsec/s rkB/s wkB/s
avgrq-sz avgqu-sz await svctm %util
/dev/sda 0.00 4433.53 0.00 379.63 0.00 37231.30 0.00
18615.65 98.07 212.70 26.69 2.63 100.00
Device: rrqm/s wrqm/s r/s w/s rsec/s wsec/s rkB/s wkB/s
avgrq-sz avgqu-sz await svctm %util
/dev/sda 0.00 4061.17 0.00 183.80 0.00 33039.00 0.00
16519.50 179.76 231.48 59.67 5.44 100.00
Today, I built and installed 2.4.20 from kernel.org with XFS patch
2.4.20-2003-01-14_00:43_UTC and Trond's NFS_ALL patch. Under a similar
test environment, the numbers are remarkably different, e.g.,
Device: rrqm/s wrqm/s r/s w/s rsec/s wsec/s rkB/s wkB/s
avgrq-sz avgqu-sz await svctm %util
/dev/sda 0.00 1215.67 0.00 356.38 0.00 11309.88 0.00
5654.94 31.74 114.67 32.16 2.80 99.90
Device: rrqm/s wrqm/s r/s w/s rsec/s wsec/s rkB/s wkB/s
avgrq-sz avgqu-sz await svctm %util
/dev/sda 0.00 1440.73 0.00 370.63 0.00 13042.82 0.00
6521.41 35.19 74.39 20.06 2.69 99.82
Device: rrqm/s wrqm/s r/s w/s rsec/s wsec/s rkB/s wkB/s
avgrq-sz avgqu-sz await svctm %util
/dev/sda 0.00 1558.33 0.00 356.70 0.00 13785.85 0.00
6892.93 38.65 59.97 16.80 2.80 99.80
The number of bytes getting written to disk is less than half that
reported above, but the await and svctm times are consistently lower
too (as is %util). Also, the await/svctm numbers in the home-brewed
kernel are much more consistent; they bounce around a *lot* more under
the SGI/Red Hat kernel.
Oddly, however, when the clients are pushing 8K reqs, hence maximizing
NFS [rw]size, iozone is reporting sequential writes at ca. 8500 kB/s,
which is pretty good for a 100Mbps link (right?).
IOW, the clients look and feel happy -- their numbers are largely
pretty good -- but the server-side numbers are much lower.
That sounds bizzare to me. What am I missing?
--Paul Heinlein <[email protected]>
-------------------------------------------------------
This sf.net email is sponsored by:ThinkGeek
Welcome to geek heaven.
http://thinkgeek.com/sf
_______________________________________________
NFS maillist - [email protected]
https://lists.sourceforge.net/lists/listinfo/nfs