Why am I not using 100% of cpu 4492

Try running only one process at a time (if possible), or put all MPI processes on a single CPU. If its CPU utilization rises over 74.2%, then the two processes are competing for a shared resource, perhaps the memory bus. Time spent waiting for memory access should lies in kernel routines, thus is system time, thus does not show up as process (user) time.

But it is odd that the process' user+system time is still below 100%. Even if the time were spent in I-O (or memory access), the complement of user time should add up to 100% (if the process is CPU bound).

Unless the scheduler is putting the process to sleep while waiting for I-O... like background NFS or network traffic. You mentioned that context switches are numerous. That's a potentially significant loss of CPU time. What other processes are running concurrently? Maybe you should drop unix runlevels to shut down X and secondary processes to see if they are stealing cycles. "init -s" will switch your workstation to single user mode.

I'd also try cranking up the priority of your process to see if that gives it a higher fraction of the CPU. "nice -20 .-program" should do it.

Also, you might run strace on the process to see if it's making system calls, perhaps to disk. Same for using vmstat to check on virtual memory use.

BTW, AFAIK all MPI implementations use polling when blocking on a send-recv. This chews up CPU cycles something fierce with busy waits, but it's the only way to achieve low latency. It can easily account for less than optimal CPU utilization. And it's unavoidable.

Randy

Dave Sill

--

"If English was good enough for Jesus Christ, it ought to be good enough for the children of Texas." -- Texas Governor Ma Ferguson (1924)