[CSC 435] Reading and MPI Thread Support

[Andrew J. Pounds]

So chapter 11 of your text is all about hybrid MPI-OpenMP programming.   
Some of you are reporting that you are seeing little to no speedup.

I am seeing something like the following...

Nodes    Threads/Node     Megaflops

1            1                         627.363474
1            2                       1420.332565
2            2                       2610.730625
2            4                       4881.539916
4            2                       4363.807638
4            4                       7121.024514

Clearly, in this small sample you can see that for the same number of 
nodes, increasing the number of threads enhances the performance.  You 
can also see that the same number of total processes (like 2 nodes and 4 
ppn or 4 nodes and 2 ppn) does not necessarily result in the same 
performance.  That is exactly what you are trying to show (and that this 
also changes based on the dimension of the problem).  You will find the 
"sweet spot" of nodes and processors for each problem size on the final.

I think what most of you are missing is that you have to do some minimal 
reading -- and note that you will most likely have to change your 
MPI_Init function to the MPI_Init_thread function call (p 268-269 of 
your text).  Since I have requested that you use only one MPI call per 
node, I recommend that you use the MPI_THREAD_FUNNELED level of support.

Keep me posted.  I have a day full of ZOOM calls tomorrow, so I am going 
be working from the house. Please let me know if you need anythin or if 
the cluster is giving your problems.


-- 
*/Andrew J. Pounds, Ph.D./*
/Professor of Chemistry and Computer Science/
/Director of the Computational Science Program/
/Mercer University, Macon, GA 31207 (478) 301-5627/
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://theochem.mercer.edu/pipermail/csc435/attachments/20240424/e09b11fe/attachment.html>