[Omp] OpenMP speedup is too good to be true?

[Monty Swaiss]

You are probably looking at a cache hit effect. Usually you do not see 
this kind of superlinear speedup from cache hits only, but if your grid 
size is large enough and memory/cache  size is small enough, it may be 
possible. Eventhough jacobi is a fairly small program, it does process 
the entire grid several times. If you are getting a small benefit from 
cache hits, this benefit is magnified depending on the number of 
iteration you specify for the solver.

What numbers do you see if you cut the grid size to 25% of what you used 
in your earlier test? I imagine you will probably see speedups less than 
"n" in magnitude.

--Monty

Open MP wrote:

>Dear all,
>
>I tried the famous Jacobi code on a 4-way Itanium/Linux box using
>Intel compiler 8.0. No any special compilation options were used
>except -openmp.
>
>The measured speedup compared to the sequential execution are 1,  2.3,
> 4.4, and 6.8 for 1,2,3,and 4 threads respectively.
>
>I expected the ideal speedup would be at most n when using n threads.
>How could it be actually much larger than n? What are the reasons
>behind these observations?
>
>Thanks in advance.
>
>An OpenMP fan
>
>_______________________________________________
>Omp mailing list
>Omp at openmp.org
>http://openmp.org/mailman/listinfo/omp_openmp.org
>
>
>  
>