I'm doing a senior project, and want to put together a folding@home style program to do something cool/useful for real people -- any thoughts on possible uses of massively distributed computing?
Especially the first three would give you something to put your teeth in because they are decidedly non-local and it would be pretty tricky to get that to work in a coarse grained parallel environment with a high cost of communication between nodes.
Most of the project that are being done right now fall in to the 'embarrassingly parallel' group (and it would be nice to see that change).
Most of the project that are being done right now fall in to the 'embarrassingly parallel' group (and it would be nice to see that change).
Applying the distributed computing approach to non-EP problems is challenging; this was my doctoral research project until I got distracted by delta compression and string matching problems. Three are three major problems:
1. Bandwidth. Many parallel algorithms require more bandwidth than the available pool of idle home computers can provide.
2. Latency. Many parallel algorithms require low latency -- say, under 1 ms -- and the speed of light thus prohibits any worldwide computing efforts. In some cases, this can be attacked by replacing algorithms (e.g., LU factorization by block QR factorization) or by using a small well-connected set of nodes to perform latency-sensitive work while most nodes do less latency-sensitive work.
3. Byzantine failures. Home PCs have a very high failure rate -- not only where permanent and transient failures are concerned, but also where returning bogus results is concerned. The best estimates I've seen indicate that bit error rates on home PCs are at least two orders of magnitude higher than those on supercomputing clusters.
I was cold-called by two VCs back in 2000, after I announced the computation of the quadrillionth bit of Pi; at that point distributed computing was a hot topic and over $50M had been thrown at companies in the area in the previous 10 years. I told them that they didn't understand the field, that they were fools for investing in it, and that the research belonged at universities for the present time... and history proved me right.
matter simulation
weather prediction
ray tracing
data mining
distributed search engine
Especially the first three would give you something to put your teeth in because they are decidedly non-local and it would be pretty tricky to get that to work in a coarse grained parallel environment with a high cost of communication between nodes.
Most of the project that are being done right now fall in to the 'embarrassingly parallel' group (and it would be nice to see that change).