Utilities for parallel computation

• dyntest: There are two versions around of the Linpack benchmarck (both can be downloaded from http://www.netlib.org ). The 100x100 and 1000x1000. They consist of a factorization (and backsubstitution) of a matrix of the given size. The linear system is solved and the megaflops rate is computed as a relation between the time spent and the operation count, which is 2 n^3/3 + 2 n^2 . The results given by both tests is very differente since normally the 100x100 test fits completely in the processor's cache while the second not. (Currently caches are 256kbytes). In addition, the 100x100 test is run in current processors in such a small time that the PC clock slice is not enough to register the event, and the reported megaflops rate is erroneoues. This version: first we loops as many times as needed in order to get a realistic estimate of the time spent, even for very low n. Second we make a sweep for n from 20 to 1000 (as in the "cpu-rate" test) so that we can see the incidence of the cache size on processor speed. (download)

• nettest: Send packets of different sizes from one node to other. Reports bandwidth and latency. (NetTest code)