This question was asked during the webcast, An Introduction to High Performance Computing: Parallel Computing Issues. The presenter, Tom Lehmann, is the Advanced Projects Manager for the Enterprise Systems Group training organization at Intel. Here is Tom's answer to the question.
You can mix and match your load balancing as you see fit. It all depends on how much effort you put into the load balancing as to whether it's actually going to get the job done any quicker. I have seen situations where a given cluster was spending so much time doing load balancing that it actually was not doing any calculations to speak of. When we added a second node to it, the thing went from one to 1.1 instead of two. So the speedup was ridiculously minor simply because it was spending all of its time doing load balancing. So whatever technique you use, static, dynamic, or some mixture, you could do that. It's just a matter of how much time are you going to spend doing the load balancing versus how much time are you spending actually doing useful work and getting to your end goal.
What other load balancing questions do users have?
