[Omp] Could you help me about this problem?

[Ruud van der Pas]

Hi David,

>    We used the OpenMP to parallel a subroutine(only a loop) of a program 
> recently. When I set different thread number, the time they took is just 
> like the below:
> 
>    1 thread:  20s
> 
>    2 threads: 11s
> 
>    3 threads: 10s
> 
>    4 threads:  7s
> 
>    5 threads: 11s
> 
>    6 threads: 19s
> 
>    8 threads: 23s
> 
>   The machine we used is a smp node(dual core) of a cluster. What I 
> can't understand is that why the time between 2 threads and 3 threads is 
> so small?

I'm afraid we need more information to say something sensible
about this, so I'm going to ask some questions first.

- I assume the above timings are the _elapsed_ time for the
   application (or just the subroutine), not the CPU time? In
   other words, how did you get these numbers?
- Was your job the only one running, or could there have been
   interference from other jobs?
- Can you reproduce it? That is, if you run the same experiments
   again, do you get comparable timings?
- To me, the surprising number is the one on 4 threads. If I
   understand it correctly, you only have 2 cores to run your
   program on. Based on that, it seems you get good scalability
   (but keep my previous question in mind) going from 1 to 2
   threads. I would expect the 4 thread number to be in the ~10s
   range (at best) though, not 7s

Kind regards,
Ruud
----------------------------------------------------------------
Senior Staff Engineer             Email: ruud.vanderpas at sun.com
Systems Group                     Phone: +31-33-4515000 (x15920)
Sun Microsystems                  Fax  : +31-33-4515001
----------------------------------------------------------------