Integrating Bulk-Data Transfer into the Aurora Distributed Shared Data System

Abstract

The Aurora distributed shared data system implements a shared-data abstraction on distributed-memory platforms, such as clusters, using abstract data types. Aurora programs are written in C++ and instantiate shared-data objects whose data-sharing behaviour can be optimized using a novel technique called scoped behaviour. Each object and each phase of the computation (i.e., use-context) can be independently optimized with per-object and per-context flexibility. Within the scoped behaviour framework, optimizations such as bulk-data transfer can be implemented and made available to the application programmer. Scoped behaviour carries semantic information regarding the specific data-sharing pattern through various layers of software. We describe how the optimizations are integrated from the uppermost application-programmer layers down to the lowest UDP-based layers of the Aurora system. A bulk-data transfer network protocol bypasses some bottlenecks associated with TCP/IP and achieves higher performance on an ATM network than either TreadMarks (distributed shared memory) or MPICH (message passing) for matrix multiplication and parallel sorting.