PGAS Model for the Implementation of Scalable Cluster Systems∗

This paper introduces an extended version of the tra-ditional Partitioned Global Address Space (PGAS) model, for the implementation of scalable cluster systems, that the HyperTransport Consortium Advanced Technology Group (ATG) is working on. Using the Simics and GEMS simu-lators, we developed a software module that approximates the behaviour of a PGAS cluster. This approach mainly provides the simplest mechanism to evaluate how much the PGAS infrastructure will affect overall the application per-formance. The aim of this work is to study the feasibility of the ATG’s PGAS model, for running applications with high memory requirements. It is not focused on measure the performance of applications, although some performance measures are given. Such a model, will let manufacturers build clusters that enable the execution of these applica-tions, in such a way that it will be impossible to run them in a single processor without adding more memory, or in a multi-processor without modifying the source code of the application. 1