A Multithreaded Message-Passing System for High-Performance Distributed Computing Applications

High Performance Distributed Computing (HPDC) applications require low-latency and high-throughput communication services and HPDC applications have diierent Quality of Service (QOS) requirements (e.g., bandwidth requirement, ow/error control algorithms, etc.). The communication services provided by traditional message-passing systems are xed and thus can not be changed to meet the requirements of diierent HPDC applications. NYNET (ATM wide area network testbed in New York state) Communication System (NCS) is a multithreaded message-passing system developed at Syracuse University that provides high-performance and exible communication services. In this paper, we overview the general architecture of NCS and present how NCS communication services are implemented. NCS point-to-point communication is exible in that users can conngure eecient point-to-point primitives by selecting suitable ow control, errror control algorithms, and communication interfaces on a per-connection basis. Furthermore, NCS architecture separates the data transfer and control transfer functions that allows the control information to be transmitted over the control connections and thus improves the perforamnce of the point-to-point communication primitives. We analyze the overhead incurred by using multithreading and compare the performance of NCS point-to-point communication with those of other message-passing systems such as p4, PVM, and MPI. Benchmarking results indicate that NCS shows comparable performance with other systems in transmitting small messages but outperforms other systems for large messages.