Comments on <Eran_Chinthaka's_core_dump/>: Optimizing MPIs for multi-cores : Thoughts

OpenMPI (Developed by Indiana) already uses optimi...

2009-06-29T00:42:16.527-04:00

OpenMPI (Developed by Indiana) already uses optimized communicaiton ( memory to memory) for in-node communication.

In most MPI implemenations(unless they are pretty old like LAM MPI) these features are available.

It is better to check how much we can gain by the hibrid model. My experiance is that with MPI, this does not buy you mutch.

Hi Steve, I agree with you about the MPI's br...

2009-06-28T23:02:52.373-04:00

Hi Steve,

I agree with you about the MPI's brittleness. But wait, MPI is/was meant to run on supercomputers and the goals were performance, scalability and portability. I don't think robustness was ever a goal, simply because it is always assumed to run on reliable hardware.
Until recently we never thought of building large super-computing clusters with commodity hardware

MPI apps can be pretty brittle against system fail...

2009-06-28T17:27:52.648-04:00

MPI apps can be pretty brittle against system failure. Our local university wont run simulations that take >72 hours as P(node failure) is too high. MapReduce has recovery built in.