Distributed Shared Memory on Network of Workstations with TCP/IP

M.E. (Reg.) Thesis, Department of Computer Science and Automation,
Indian Institute of Science, Bangalore, India, January 1997.

Authors

1. Biren Gandhi, Department of Computer Science and Automation

Abstract

Shared Memory Multiprocessors are relatively easier to program but have lower scalability with increased number of processors. Distributed Memory message-passing multicomputers scale well but are cumbersome to program. Distributed Shared Memory (DSM) systems attempt to combine the advantages of both approaches. Software DSM systems do not require any special hardware and hence can be implemented on any platform. Commonality of workstation networks with very high total processing power and higher costs of multiprocessor systems motivate the implementation of DSM on Network Of Workstations (NOW) as a low-cost, high-performance alternative.

We have implemented DSMNOW, Distributed Shared Memory on Network Of Workstations over the TCP/IP protocol on a network of UltraSparc workstations connected via 10 Mbps Ethernet. This implementation leads to reduction in communication overhead and increased ease of programming by careful selection of memory consistency model and other DSM policies. The salient features of our implementation are:

• user level library without OS or compiler support
• lazy release consistency memory model
• hybrid coherence policy
• multiple-writer protocol with eager diff creation

Lastly, the performance of DSMNOW is measured by running 3 benchmark programs in DSMNOW environment.

Download

Full Text