High-Performance Cluster Computing Using Scalable Coherent Interface

In this paper we discuss our experience with highperformance cluster computing using Scalable Coherent Interface (SCI), an IEEE communication standard. We present detailed measurements of basic SCI transactions on a cluster of 8 UltraSparcs connected via SBus-2 SCI adapter cards and an SCI switch. Our experiments show that the simple remote load and store transactions supported by SCI allow for a very low latency implementation of data transfers, resulting in much lower latencies than other cluster of workstations using networks such as Myrinet, and parallel machines such as the Meiko CS-2. We have implemented Active Messages (a low-latency communication substrate) and Split-C (a simple global address-space language) on this cluster. Supporting Active Message requires implementing a remote queue. We discuss several implementation alternatives for remote queues and evaluate these under micro-benchmarks as well as under full Split-C applications. Our performance measurements show that ma..