Platform for Principled Experimentation of Hard Computational Problems

摘要

Our research in the area of computational complexity of combinatorial problems involves intensive use of computational resources. Traditionally, researchers with such compute intensive needs use supercomputers in a time and node sharing fashion. There are some problems with this approach: (a) Supercomputers tend to be very expensive to acquire (millions of dollars) and to maintain. Also the competition used to be quite weak. (b) Because supercomputers are expensive, parts required to repair them are expensive also, thus the maintenance costs are high. (c) A number of researchers have access to supercomputers competing for CPU time (and occasionally other resources), thus quite often one has to wait weeks until his/her tasks are run. (d) It is expensive to run something on a supercomputer (since supercomputers are themselves expensive), thus quite big grants are required to do anything serious.