computational workstation: balancing ram vs cpu

[antineutrino]

Hello all,

I'm trying to spec a new workstation to be used for heavy computational tasks (physics modeling, with experiment runs potentially running into a week or more).

I don't really know how to benchmark the runs to figure out if they are more dependent on CPU or RAM, and thus buy the right balance of processor and ram in a computer.

For example, the battery of tests used to validate the software installation on a given computer takes, in practice, something around 1000 processor minutes (according to the manual for the software).

Does anyone have an explanation of evaluating CPU vs RAM that would work for someone that has a rudimentary knowledge of computer hardware? I have a fixed budget, so I don't want to waste money by buying a combination of CPU and RAM that isn't optimal for my use.

Thanks

 

[mfenn]

Basically, with RAM the rule of thumb is that if you have enough to fit your entire task into memory, then any more beyond that isn't going to improve performance. I would imagine that you can do a rough back-of-the-envelope calculation of how much memory that your models are likely to use, add 20-30% as a safety margin, and then buy enough RAM to fit your model. Spend the rest on as fast of a CPU as you can afford.

 

[antineutrino]

Basically, with RAM the rule of thumb is that if you have enough to fit your entire task into memory, then any more beyond that isn't going to improve performance. I would imagine that you can do a rough back-of-the-envelope calculation of how much memory that your models are likely to use, add 20-30% as a safety margin, and then buy enough RAM to fit your model. Spend the rest on as fast of a CPU as you can afford.

Okay, that makes sense. So I need enough RAM to fit the task in memory PER CPU though (highly parallel computations). What if I have lookup tables on the disk, how does that calculate into this? BTW, the lookup tables range up to about 20 gigs.

I appreciate the help, thanks!