Fat-tree for local area multiprocessors

The local area multiprocessor (LAMP) is a newly-feasible approach for achieving high-performance low-cost parallel computing. LAMP allows a large number of machines distributed over a LAN-size area to operate as a distributed-shared-memory multiprocessor. Latency is kept low by using the scalable coherent interface and by caching remote data. SCI keeps the caches consistent and provides a high performance multiprocessor “bus” using only point-to-point physical connections (links), which can operate at far higher speeds than any bus. Large numbers of links may be used concurrently for extremely high system throughput, interconnected by a packet switch using any of a wide range of topologies and technologies. This paper examines the use of a fat-tree topology for this (possibly distributed) switch. Simulation results are presented to show the latency, throughput, buffer requirements, and the effect of cable length. Link clock speeds assumed are 250 MHz (GaAs or biCMOS chips) and 32 MHz (CMOS)