[pjscott]

There's only so much parallelism you can exploit, especially with network lags adding overhead.

http://en.wikipedia.org/wiki/Amdahl%27s_law

[zaptheimpaler]

Thats a pretty trivial bound though. It basically deals with some portion "p" of the program that can be parallelized and tells you how that would affect performance given that the rest (1-p) can't be parallelized. "p" would probably be a huge chunk of the program in most cases, say a DB lookup if that is the biggest bottleneck.

[jff]

I'm sure the supercomputing community would love to hear how communication costs and Amdahl's law are pretty trivial.