MATLAB Distributed Computing Server - dedicated cluster vs. pool of PCs

good afternoon, I was looking at the MATLAB Distributed Computing Server to run multiple jobs/tasks on multiple nodes/workers. I am a bit confused whether I should use a standard pool of PCs connected to same network or dedicated cluster. Your response on this matter will be very much appreciated.

Best regards

Pasquale

Answer 1

[Walter Roberson]

There are arguments on both sides. Dedicated clusters tend to support greater density and lower power usage, and to have interconnects that are optimized, and to support higher end CPUs. Ordinary PCs tend to be inexpensive especially if you can get them used.

As a very rough approximation, the PCs can do well when it is a matter of dividing up the problem space into many independent pieces where the performance per core varies by less than the cost of faster cores. Ummm, I kind of know what I mean. It might be cheaper to buy two 2 gigahertz systems than one four gigahertz system and if you have unlimited electricity then if you later add one of those 4 gigahertz systems you might as well keep the two slower systems as well if they are helping chew through problem independently. But if you have limited electricity then it could sometimes even make sense to replace two 2 gigahertz system with a single 3 gigahertz system - you have to look at the total running costs per unit of progress.

Now on the other hand, if you have a computation that needs a lot of communication, where the different subcomputations influence each other a lot then the closer integration of dedicated clusters can make a difference. Weather forecasting is an example.

One difficulty that we used to run into with working with parallel systems was that it was difficult and expensive to find switches that could sustain high throughput. These days you can get sub $100 wireless access points that have 4 to 8 gigabit ports added for convenience, but I would not want to trust that they could sustain even half a gigabit aggregate, not without testing.