Distributed shared variables for user-level fine grained multithreading

Abstract

Scalable networks of workstations are a cost effective and hence attractive high performance computing solution. Distributed computing is commonly based on the message passing protocol. This is considered to be a complex and consequently error prone environment for the programmer. Distributed shared memory (DSM) is a software abstraction that offers the view of a shared memory address space on a series of distributed machines, thus simplifying the programming task. Distributed shared variables (DSV) is a special type of DSM where each shared variable is treated as a separate entity. As a consequence, DSV systems are suited for fine grained shared memory programming. The key problem in developing an efficient DSV system is to reduce the network latency and the amount of communication required to maintain the shared variables consistency while allowing useful computation to occur at variable synchronisation points. We thus present COSHAVA, a high-performance user-level DSV system that offers a simple programming interface. COSHAVA is built on top of a user-level multithreading library, that allows computation to overlap with comm1mication and hence hides network latency. The consistency-related communication is minimised by using an efficient implementation of the lazy release consistency model. COSHAVA accomplishes all the underlying communication using synchronous channels built over the traditional TCP IP protocol. In addition, we show the versatility of COSHAVA by applying it to implement a distributed semaphore system.