[Insight-users] Profiling ParallelSparseFieldLevelSetImageFilter

[Brady McCary]

insight-users,

I have done some superficial profiling of a level set method using
ParallelSparseFIeldLevelSetImageFilter. The results are as follows:

1 Thread(s)
real	12m0.678s

2 Thread(s)
real	20m51.607s

3 Thread(s)
real	18m27.440s

4 Thread(s)
real	17m2.819s

5 Thread(s)
real	15m57.489s

All runs were exactly the same except for the NumberOfThreads argument
to the filter. As you can see, I am experiencing a penalty for
multithreading. I have not tested more than 5 simultaneous threads,
but I will test 6, 7, and 8 threads this week when the hardware is
available to me. If the trend continues, then probably around 7 or 8
threads I will benefit from multithreading.

However, this is contrary to my reckoning, and my computation should
scale (mostly) linearly with the number of threads. I would like to
discover the culprit. Would someone familiar with this filter
recommend a general course of action?

For posterity, some keywords: thread, profile, gprof.

Thanks,

Brady