[r2d2]

Hi,
I am running a the parallel version (with LAM) for some modification I´ve done on oodles and what I observe is that the processes use only about half the CPU on each node. Am I doing something wrong or that´s the way it should work? Does it have anything to do with LAM?
Cheers,
Radu

[henry]

I assume your node interconnect is not very fast in which case you are network bandwidth limited and the CPUs are waiting for data. You might want to try playing with the scheduledTransfer and floatTransfer parameters in

OptimisationSwitches
{
fileModificationSkew 10;
scheduledTransfer 1;
floatTransfer 0;
nProcsSimpleSum 16;
}

which is in .OpenFOAM-1.1/controlDict

In particular try setting

floatTransfer 1;

and see if which setting (1 or 0) for scheduledTransfer gives best performance.

[r2d2]

OK, thanks for your answer. In the meantime, whilst I play with the parameters, can you please tell me what do these parameters mean? I mean all of the OptimisationSwitches...
And about the bandwidth of our network...will ask the admin.
Cheers,
Radu

[henry]

Here are the definitions and descriptions from the FoamX configuration of the controlDict file from the 1.2 release:


OptimisationSwitches
{
type dictionary;

entries
{
fileModificationSkew
{
type label;
description "Time delay of automatic re-reading of case files to compensate for time difference between networked machines";
}

scheduledTransfer
{
type label;
description "Use data scheduling during processor-processor transfers rather than asynchronous buffered transfers";
}

floatTransfer
{
type label;
description "Convert doubles into floats for processor-processor field transfer to reduce bandwidth (transfers the mean value as a double and renormalises)";
}

nProcsSimpleSum
{
type label;
description "Number of processors at the global-sum method changes from flat to hierachical";
}
}
}