I'm doing some benchmarking on a dual 1.8GHz Opteron system
with 4GB of RAM running SuSE 8.2 Pro BETA 9 - and mostly it
looks nice, except for the occasional strange hickups. For example,
I compiled the standard LINPACK Fortran benchmark program. If I
run the same program twice in a row then I sometimes see a *long*
delay in running time. A run with "time" shows that the delay is
not in "user" time but in system:
[1] opteria:~/src/linpack> time out/opteria/linpack.g77
...
xtimes are reported for matrices of order 1000
factor solve total mflops unit ratio
times for array with leading dimension of1001
2.450E+00 0.000E+00 2.450E+00 2.729E+02 7.328E-03 4.375E+01
end of tests -- this version dated 10/12/92
2.520u 0.020s 0:02.53 100.3% 0+0k 0+0io 158pf+0w
---> Ca 273 MFLOPS.
[0] opteria:~/src/linpack> time out/opteria/linpack.g77
Please send the results of this run to:
...
xtimes are reported for matrices of order 1000
factor solve total mflops unit ratio
times for array with leading dimension of1001
2.697E+01 8.000E-02 2.705E+01 2.472E+01 8.091E-02 4.830E+02
end of tests -- this version dated 10/12/92
2.680u 25.370s 0:28.04 100.0% 0+0k 0+0io 158pf+0w
---> Ca 25 MFLOPS, and a whopping 25.37 seconds spent in system-space.
Any suggestions as to why I'm seeing these delays? The machine
is more or less idling...
[0] opteria:~/src/linpack> uname -a
Linux opteria 2.4.21-60-smp #1 SMP Tue Sep 2 11:34:54 UTC 2003 x86_64 unknown
unknown GNU/Linux
I've tried the NUMA kernel, but it doesn't make any difference.
--
Peter Eriksson
On Fri, Sep 19, 2003 at 06:15:05PM +0200, Peter Eriksson wrote:
I'm doing some benchmarking on a dual 1.8GHz Opteron system with 4GB of RAM running SuSE 8.2 Pro BETA 9 - and mostly it looks nice, except for the occasional strange hickups. For example, I compiled the standard LINPACK Fortran benchmark program. If I run the same program twice in a row then I sometimes see a *long* delay in running time. A run with "time" shows that the delay is not in "user" time but in system:
You could profile it. oprofile is included and it will profile both your application and the kernel.
I've tried the NUMA kernel, but it doesn't make any difference.
NUMA kernel is identical to SMP kernel currently. -Andi
participants (2)
-
Andi Kleen
-
Peter Eriksson