It also depends on how you plan to use it. If you want to run many small to medium problems with many sets of input parameters or structure variations, a group or reasonably fast computers connected with 100Mb/s ethernet can still give very good performance since the problems are essentially independent and don't need to communicate much. I use Apple's Xgrid to do this on a network of G5 powerMacs with quite good results.
If, however, you are trying to run one large problem over many processors, high bandwidth interconnect like Myrinet is probably needed. On Jun 21, 2007, at 11:32 AM, Steven G. Johnson wrote: > On Wed, 20 Jun 2007, Francesco Riboli wrote: >> we are planning to buy cluster of computers to perform extensive >> numerical simulations. In order to choose the best system for >> meep, we >> need to know for which type of architecture Meep is optimized. It is >> better to have one node with multiprocessors and multicore or to >> distribute the calculus on diverse node of the cluster? Many Thanks > > It's important to have a fast interconnect between the processors. We > normally either run Meep on a shared-memory system (e.g. 4-core Intel > processors) or on a high-end supercomputer with a fast interconnect > (e.g. > Infiniband or Myrinet). For these systems, we've gotten pretty good > (within a factor of two of linear) speedup, going up to a couple > hundred > processors for large problems. > > It's possible that gigabit ethernet may be sufficient for a large > enough > problem, but I can't recall benchmarking this myself. Perhaps someone > else on here has some numbers. > > Steven > > _______________________________________________ > meep-discuss mailing list > [email protected] > http://ab-initio.mit.edu/cgi-bin/mailman/listinfo/meep-discuss _______________________________________________ meep-discuss mailing list [email protected] http://ab-initio.mit.edu/cgi-bin/mailman/listinfo/meep-discuss
