Barak.Pearlmutter@CMU-CS-G.ARPA (01/31/84)
The Boltzmann group here at CMU is interested in a simulation problem which can be sped up considerably by running identical copies of the simulator on multiple machines, and then combining the results collected and behaving appropriately at relatively infrequent intervals. The volume of data to be combined is small compared to the total size of the simulation. It would seem that an ideal (i.e. cheap and off the shelf) vehicle for doing such a simulation would be an IBM PC with lots of boards plugged into it, each board containing a local processor, a good hunk of local memory, and some way of communicating with the native 8088. For this application lots of them have to be running at the same time, so we are not interested in boards that take over the entire machine when they run. Nor are we really concerned with good OS support; we're willing to hand code stuff to load the simulator into the slaves and examine the results ourselves. We just want the little guys to crunch. If anyone has done similar things with IBM PCs (or other off the shelf hardware) we'd love to know about problems you encountered, and if you have information about such boards, especially concerning their speed (we can get away with fixed point math), cost, amount of local memory and communication channel configuration, we'd appreciate anything you can tell us. Multiple thanks in advance... Barak Pearlmutter Pearlmutter@CMU-CS-C.arpa Department of Computer Science Carnegie-Mellon University (412) 578-2582 Schenley Park Pittsburgh, PA 15213