"The Internet is not a CPU backbone. So here's the algorithm. If the instruction density, that is the number of instructions you execute per byte of network traffic, the CPU time per byte of network traffic - if it's greater than 100.000 - ship it. And if the CPU at the other end is free, then it's a positive ROI. If you have less than 100,000 instructions per byte, it's a negative ROI. So, most of the people I work with, Renderman, is a little under 100k per byte, ray tracing kind of things. Astronomers I work with, they're looking for galaxies, doing point spread functions and doing curve fitting, and basically doing vision on pixels - they're at the 10,000 instructions per byte density. And MPI, if you at it, is way way down on the instructions per byte, which is to say 1,000 instructions per byte, 10,000 instructions per byte for a typical MPI program. Fundamentally this is a way of saying, if you have a computation, when should you do it in the cluster and when should you try and ship it out into the wider Internet, and the answer is this. Now of course, this algorithm and this equation changes depending of the cost of processing and the cost of network." - Jim Gray, Distributed Computing Economics...
Previous Topic
