[Boost-users] hybrid parallelism

Subject: [Boost-users] hybrid parallelism
From: Hicham Mouline (hicham_at_[hidden])
Date: 2010-11-03 18:43:11

Hello,

Following a previous thread that asked about how to parallelize a large
number of calculations, I took this summary:

 

I am thinking of choosing a simple model whereby:

There are M computers (possibly heterogeneous). I will stick to 1 process
per computer.

Each process will have N threads. Giving a total M*N "execution units".

I will take a simple solution in that the M and N are fixed though
determined at runtime.

 

So there is this pool of M*N exec units and I give them 100 000 tasks to do.

Those tasks get scheduled in the same process/computer, they share their
memory.

Each of the computers gets a duplicate of the memory used as inputs to the
tasks.

 

For thread pool, TBB, boost::asio were suggested. I believe there is also
threadpool.sourceforge.net

For the cross-computer communication, boost::mpi.

 

Is this something boost::mpi + mpi can help with?

 

Regards,