Predicting the performance of distributed virtual shared-memory applications

The use of networks of workstations for parallel computing is becoming increasingly common. Networks of workstations are attractive for a large class of parallel applications that can tolerate the higher network latencies and lower bandwidth associated with commodity networks. Several software packages, such as TreadMarks™, have been developed to provide a common view of global memory, allowing many shared-memory parallel applications to be easily ported to networks of workstations. This paper investigates in detail the performance of several TreadMarks-based shared-memory applications on a modern network of workstations, identifying the extent to which different system components affect the efficiency of these applications. Then a performance model for such applications is developed and used to evaluate the impact future changes in technology are likely to have on performance. The results of the model indicate that current systems are limited in their performance by communications and software overhead for supporting the distributed virtual shared memory, rather than hardware delays.