Shared Memory Computing on Networks of Workstations

Willy Zwaenepoel

Department of Computer Science
Rice University

By building on existing infrastructure, networks of workstations (NOWs) provide a low-cost, low-risk entry into the parallel computing arena. Furthermore, using a shared memory programming model, existing sequential codes can be parallelized with much less programmer effort than when using message passing. The combination of the two, shared memory programming on a network of workstations, thus provides excellent leverage both for hardware and software investments. It also provides some measure of portability between SMPs and NOWs.

We have developed a runtime package, called TreadMarks, that provides a shared memory image to processes executing on different workstations. The package is relatively portable, and runs on most common Unix platforms (DEC, HP, IBM, Intel, SGI, and SUN). No kernel modifications or special privileges are required to run TreadMarks programs. C, C++, and Fortran are supported, using standard compilers and linkers.

While the appeal of shared memory programming has been well known for some time, early software implementations have suffered from poor performance due to excessive communication. We have developed a number of techniques to address the communication problem. In this talk I will discuss the two principal techniques: lazy release consistency and multiple-writer protocols, and contrast them to the sequential consistency and single-writer protocols used in conventional systems.

I will demonstrate the programmability and efficiency of TreadMarks by discussing the parallelization of FASTLINK, a popular package for genetic linkage analysis, the computational step in disease gene location. I focus on this application because for a number of reasons it is hard to do in message passing and at the same time it stretches the performance of distributed shared memory systems. Nevertheless, we have been able to achieve good performance for a large number of datasets. The resulting parallel FASTLINK package is currently in use at a number of sites, both on SMPs and on NOWs. It was recently used in the discovery of a linkage for Parkinson's disease by researchers at NIH.

Back to LESS

Last modified: February 14, 1997
Mike Dahlin