SHF:Small:Fine-grain Tasking and Virtual memory for Massively Parallel Computing

Massachusetts Institute Of Technology, Cambridge MA

Investigators

Abstract

In order to address the performance and programmability challenges of massively parallel systems new innovative approaches are needed. This research continues earlier work on the Fresh Breeze model of computation for which basic simulation results have demonstrated applicability to linear algebra kernels as well as the Graph500 challenge application with competitive performance. The merits of the programming model developed in the proposed research are demonstrated, and evaluated for usability and performance, using a new distributed simulation tool based on packet communication architecture. This project proposes innovative combination of a general tasking model, a proposal for global virtual memory, and the matching system architecture. The innovations are expected to serve as the basis for systems that are more energy-efficient and resilient. The form and structure of computer programs for massively parallel computation are studied and implementation challenges analyzed and resolved. The three components of the project concern: general support for the major forms of parallelism found in sound and well-structured concurrent programs; structured means for expressing transactions on shared data such as data bases; and demonstration of a global virtual memory based on use of tree structures to represent all data objects. This is expected to reduce development effort and increase performance of software needed to address current and future scientific, environmental and social problems.