Directed Point: A communication subsystem for commodity supercomputing with Gigabit Ethernet

Abstract

In this paper, we study the practical issues on the design of a new communication subsystem, named Directed Point (DP), on a server cluster with Gigabit Ethernet connection, with the goals of achieving high performance and good programmability. Our design exploits the gigabit network architecture and the operating system characteristics. We propose a realistic communication model which can be used to assess various design tradeoffs and to calibrate the performance results. Testing shows that DP communication subsystem can achieve a 16.3 μs single-trip latency and 79.5MBps bandwidth. To achieve good programmability, we propose an abstraction model that allows all inter-process communication patterns to be easily coded using the provided API. The API preserves the syntax and semantics of traditional UNIX I/O operations, making the proposed communication subsystem easy to use without long learning period. © 2002 Elsevier Science B.V. All rights reserved.