| DATE: | Thursday, December 7, 1995 |
|---|---|
| TIME: | 12:30pm |
| WHERE: | 111 Cummington St. Room MCS 135 |
In this talk, I'll review latency and distributed shared memory systems related research. I'll also discuss a simulated shared memory system that I have implemented and present some preliminary results.
Hardware base for distributed systems has changed dramatically, both in terms of speed and reliability. These advances in technology have increased the attractiveness of parallel computing on workstation clusters used as multicomputers. The scalability and performance of parallel applications on multicomputers depend to a great extent on communication performance. Poor communication performance not only limits the granularity of programs that can run successfully, but also degrades the performance of coarse-grain programs and limits the scalability.
Distributed shared memory (DSM) on a network of workstations provides the abstraction of a shared address space available to all the processes of the same parallel program on different workstations connected via a high speed network. Relaxed consistency models distinguish between data and synchronization accesses. These models guarantee that memory will be consistent at synchronization points and all the synchronization accesses are sequentially consistent.