Date: Sat, 19 Jul 1997 07:39:51 -0700 From: Fred Gilham <gilham@csl.sri.com> To: hackers@FreeBSD.ORG Subject: NFS wierdness? Message-ID: <199707191439.HAA12910@japonica.csl.sri.com>
next in thread | raw e-mail | index | archive | help
I have this user who runs programs that generate lots of output to
stdout. The output sometimes contains very long `lines'---long
stretches of output without a newline. The programs are also supposed
to be very CPU intensive.
Most of the time he invokes them like this:
user-program >> output
Unfortunately when he does this on a FreeBSD box where the directory
he's writing to is NFS mounted, he gets almost no CPU time; the
process spends most of its time in the `nfsfsy' state (according to
top).
Here's a program that basically simulates the user's programs:
----------------------------------------
#include <stdio.h>
int
main(int argc, char *argv[])
{
long rnum = rand();
long i;
while(1) {
for(i = 0; i < 10000; rnum = rand(), i++) {
printf("%ld ", rnum);
}
fflush(stdout);
/* Soak up some CPU time. */
for(i = 0; i < 1000000; i++)
rnum *= rnum;
}
}
----------------------------------------
I called the program `tester' for FreeBSD and `tester.sol' for the
ultrasparc.
Here's a top output line for a FreeBSD-2.2.2 box running on a Pentium
Pro 200:
16639 gilham -5 0 148K 300K nfsfsy 0:10 16.18% 15.49% tester
Here's one on an Ultrasparc 167MhZ:
6755 gilham -25 0 736K 584K run 0:42 36.86% 87.84% tester.sol
This gives an idea of the problem; usually the user sees 98% cpu
utilization on an ultrasparc and around 2% on a FreeBSD box.
Both of these boxes are on the same network segment as the directory
they're writing to. They were invoked as follows:
tester >> dummy
and
tester.sol >> dummy
Another symptom of the problem is that it's not possible to get any
output from
tail -f dummy
on the same box that's running on tester program under FreeBSD.
However, solaris has no problems.
Just for comparison we also tried invoking the program as
tester > dummy
This gave somewhat better results:
16792 gilham 47 0 148K 300K RUN 0:21 27.21% 26.66% tester
but the program still spent a lot of time in the `nfsaio' state.
With the user's program he saw a similar improvement---from 2% to 4%.
But still not very useful.
Here's the solaris top line by comparison:
6772 gilham -25 0 736K 584K run 1:00 52.97% 94.41% tester.sol
The bottom line is that the user's program runs really slowly under
FreeBSD when he tries to save the output to an nfs-mounted directory.
Does anyone have any suggestions as to what might be going on?
-Fred Gilham gilham@csl.sri.com
Want to link to this message? Use this URL: <https://mail-archive.FreeBSD.org/cgi/mid.cgi?199707191439.HAA12910>