OpenseesMP: mumps & drift recorder

[jamesgu]

I cannot find much documentation for openseesmp. I am getting errors but could not locate them.
1) What is the function of Mumps? Does it split the matrix or system to multiple processors automatically? If I can control the run within each processor, shall I switch to others, such as Umfpack?
2) The document for "recorder drift..." says for "not work in OpenSeesSP". But it seems working for me on SP. Does that relate to openseesMP? Can anybody suggest an alternative way to record the drift (peak drift only)?
Thank you,
James

[fmk]

Mumps is a parallel solver .. use with MP if using MP to do a large model (it has some problems sometimes .. bug in the mumps code that is released, but it usually works)

it will work as long as the 2 nodes in the recorder, exist on the same processor .. this is not always gauranteed and it will depend on the model and number of partitions you specify .. as such we say it does not work (you got lucky) .. it is not another constraint i want to put in the partitioning (as most record drifts between floors and constrain nodes in a rigid diaphragm on a floor, so if added the constarint that all nodes in the recorder had to exist on same processor i would get no parallel processing)

[jamesgu]

Thank you very much, fmk.

1) I could not make mumps work well. It is better to run MP based on multiple single processors.
2) It is better to record the output every individual steps, rather than the whole record from"recorder drift".
3) The "nodeReaction" does not appear to work at all. Does anybody have a way to record the reaction at each time step?

[fmk]

what are you doing with MP .. multiple independent analysis or 1 large model.

[jamesgu]

Thank you, Frank. I think that I did not understand the intention in your previous post.
I am doing parameter study with multiple independent analysis (time-history analysis with many iterations, MP), not a large model.

[fmk]

you cannot use Mumps (nor do you need to for small models) in this case.