Hi, can someone tell me how to setup the situation where I can use 1 core as a master control machine and utilise both cores on a remote machine, BUT, instead of maxing out the cores by saying I have 2 processors on it, say I have 1 processor on it, but to max out both cores during the solve?

This is that rationale, I have found that if I use distributed MPICH and select a machine this has 2 cores, it actually send 2 packets of info to that machine, the cores process the current step, send the info back and so fourth. There is a lot of communication over the network. If however I could somehow setup the run to send 1 packet of info to the remote machine, have that machine divide it up and run it on the two cores and then send the info back, this would be much fatser. How can this be done?

I have a 100 MB network and running my main machine on 2 cores (it is a 4 core CPU) and 5 remote machines on 2 cores uses upto 65% of available network overhead using Distributed MPICH. I think it would be much faster if I had to only talk to each machine once, instead of twice ( once per core )

Thanks Justin

Hi,

You cannot do what you propose.

I think you will find that for small clusters (less than 8 or 16 processors) the network does not affect the simulation speed greatly. Even if you could do what you propose, I would estimate the speedup resulting from it at around 1%-2% so it is not really worth it.

Glenn Horrocks