Hi,

I'm an IT person helping a analog/ms design team (I am not a designer.) I am building up a compute system for the team and am looking to get the most out of the software. I get that because designers are waiting for their runs to complete I need the fastest single thread simulations possible with the largest caches. To get this, I need to manage the allocation of simulations to cores in the machine. This is Cadence sw I am running on RHEL and have looked at all the things it can do. None of them seem able to get the most out of the machine for the designers.

This has led to think about writing a little scheduler that would run as a shim that would have some scheduling policy, run jobs on specific cores and stall things so that there are as few context switches as possible. This would mean understanding how many jobs are involved in this task (prioritize single runs over monte carlo sets) and what else is already running as well as what else is running on the machine.

Now a few questions:

Does this sound like a reasonable plan? I think it could make a 5-10% decrease in job completion times.

If so, are there any people who would be willing to answer questions as I dig through the ins and outs of how Cadence runs sims as well as kibitz on policy ideas?


Another question is more general. I am an IT consultant. It seems like there is very little specialized IT support for analog/mixed signal simulation work. Do people think that they might hire a consultant with these specific skills?

thanks in advance,
jerry