Very Computer

Is there an unix system command that can be used to check out how many
processors are available on a supercomputer at a certain time?

Thanks

sizer -p

man sizer for more


| Is there an unix system command that can be used to check out how many
| processors are available on a supercomputer at a certain time?
|
| Thanks
|
|

Depends on how you define "available".  As someone else noted, you
can use "sizer -p" to get the number.  You can use psrinfo(1) to
the state of each CPU.  If online counts as available, this may
be more useful to you than sizer.  If you want available to mean
"practically idle" that gets harder.

Quote:

>Thanks

--

The "practically idle" situation is what I am really concerning on. I am
running a parallel Fortan code (using openMP) on a public supercomputer.
Sometimes the performance of the parallel code is even much worse than the
sequential code if the computer is busy.  So I am wondering if I can use
some Unix system command to detect how many CPUs are really available, then
I can let the parallel code to run with the maximum number of CPUs it can
be.

Thanks

"Alan Rollow - Dr. File System's Home for Wayward Inodes."

What kind of supercomputer do you mean? I'd say, that there is no
standard command to get this information on every machine that is called
a supercomputer by someone.

Usually such systems are used with the help of some queueing system,
such as RMS, LSF, NQS, PBS,... to avoid exactly the problem you
experienced: system overload. One can find out the numer of idle
processors with the tools provided by such systems. Moreover, one can
can request as much processors as one needs for a certain job, and the
queueing system starts the job only, if there are sufficient processors
idle.

HTH,
Ralf

There are a number of ways to look at the current load on each
CPU, though vmstat and iostat aren't among them.  There is a
program called cpuinfo that is part of the one of the programming
example subsets, that displays the per-CPU data graphically.  Monitor
will display it via Curses.  Collect, mostly likely collects CPU
data per-CPU, so it probably displays it as well.  I have a recollection
of a dbx(1) command that can be used on the kernel to print the per-
CPU stats.  The tuning guide might offer a hint.

The underlying information used by these problems is available using
either table(2) or getsysinfo(2) (I don't recall which).  The per-CPU
load data is stored as the number of soft clock ticks spent in each
of the different modes since the system booted; user mode, nice,
kernel, wait and idle.  A custom program to collect it and display
it in some desired format be non-trivial, but not hard.  Especially
since the source to cpuinfo is part of the subset.  The Monitor
distribution also include sources.

Quote:>> >Thanks

--