Re: Process time accounting problem on two CPU machines. Resolved?

[Lew Randerson <LRanderson PPPL Gov>]

Following advice from the contributors, I rebuilt the kernel with
"Enhanced Real Time Clock Support" turned off and the problem
has been resolved.

People who have been living with this problem and not fixing it
should be aware that I saw a couple of other problems before
the fix. (1) Reported cpu usage via 'top' was never more that
50%. (2) The wall clock elapsed time on cpu intensive jobs
was twice what It should have been. These problems have
also gone away with the fix.

Thanks, --Lew



At 05:17 PM 6/28/00 -0400, you wrote:

> Around last October, Lars Hjorth reported a process time accounting problem to
> axp-list. Was that resolved?
>
> We are running kernel 2.2.16 on several UP2000 Dual Alpha machines and
> are seeing the same problem. i.e. There is no reported cpu usage on the
> second cpu even though we know it is being used. For each process,
> /proc/<pid>/cpu always shows cpu1 with no time. /proc/cpuinfo does
> show 2 active cpus. The machines were deliverd with 2.2.14 and had
> the problem with that kernel also.
>
>
>    # /usr/bin/uptime
>    3:48pm up 31 min, 5 users, load average: 4.08, 3.30, 1.93
>
>    # /bin/grep cpu /proc/stat /proc/{1209,1211}/cpu
>    /proc/stat:cpu 656276 0 97507 3080381
>    /proc/stat:cpu0 656276 0 97507 1163299
>    /proc/stat:cpu1 0 0 0 1917082
>    /proc/1209/cpu:cpu 112750 436
>    /proc/1209/cpu:cpu0 112750 436
>    /proc/1209/cpu:cpu1 0 0
>    /proc/1211/cpu:cpu 104966 410
>    /proc/1211/cpu:cpu0 104966 410
>    /proc/1211/cpu:cpu1 0 0
>
> Note in /proc/stat the idle time for cpu1 is essentially the same as the
> system uptime. So cpu1 is reporting NO user time. My first pass at the
> Linux source showed nothing glaringly wrong in kernel/sched.c or
> alpha/kernel/smp.c..
>
>
> Note: All utilities show CPU1 is not used including xosview and procinfo.
> Reported CPU usage is accurate when the machine is booted with nosmp
> and therefore only one cpu is active.
>
> Thanks for any info, --Lew
>
> Lew Randerson
> Computational Plasma Physics Group
> Princeton Plasma Physics Lab, Princeton University
> P.O. Box 451
> Princeton, NJ 08543
>
> lranderson@pppl.gov
> 1 609 243-3134 (voice)
> 1 609 254-3086 (fax)
>
>
>
>
>
>
> --
> To unsubscribe: send e-mail to axp-list-request@redhat.com with
> 'unsubscribe' as the subject.  Do not send it to axp-list@redhat.com