On 08/19/2011 02:15 PM, David W Noon wrote:
My experience with OpenMP is that it is difficult to write a loop body
large enough that context switching does not overwhelm the benefits of
parallelism.
Hmmm.
If you do a multiplication of a 100*100 Matrix you could spawn 10000
threads and this will result in a huge switching overhead.
But if you have 10 cores and you aggregate the 10000 tasks in 10 groups
of 1000 calculations each, spawn 10 threads and have each go through a
loop of calculating 1000 cells, I gather that (in a perfect world) no
task switching overhead at all would be necessary (but at the beginning
and the end of the complete calculation).
If in Prism you do something like (pseudo-code draft):
------------------------------------------------------------------------
m := 100;
n := 100
for parallel ij := 0 to m*n-1 do begin
i := ij mod m;
j := ij div m;
calccell (i,j);
end;
------------------------------------------------------------------------
I understand that Prism (or rather .NET) on a 10 core machine
automatically would create 10 threads each doing 1000 cells.
-Michael
_______________________________________________
fpc-devel maillist - [email protected]
http://lists.freepascal.org/mailman/listinfo/fpc-devel
