The WARPmemory prototype: design and performance

High performance multiprocessors generally require a very complex underlying hardware. This applies notably to shared memory systems, where the communication grain is fine and hence low latencies are critical. We describe the prototype of a multiprocessor network which is based on low-cost standard technology. We point out the main difficulties encountered during the design, and we demonstrate our solutions. While accepting a number of speed-versus-cost trade-offs, the resulting system remains within a close range of an optimal design.