Implementing Active Messages and Split-C for SCI Clusters and Some Architectural Implications

In this paper we discuss our preliminary experience with implementing active messages and Split-C on a cluster of workstations connected via Scalable Coherent Interface (SCI). Our experiments were conducted on a cluster of UltraSparcs connected via SBus-2 SCI adapter cards and an SCI switch. We present performance measurements and discuss several avenues for optimizations. As our experiments show, SCI offers much lower latencies than other cluster of workstation architectures, including a Myrinet based cluster, and other parallel machines like the Meiko CS-2. The simple remote load and store transactions supported by SCI allow for a very low latency implementation of the data transfer. However, active messages additionally require a transfer of control since the receiving processor has to execute the active message handler. Unfortunately, remote load and store transactions are not sufficient to efficiently implement this transfer of control. We propose hardware support for a remote enq..