Abstract

The bag of tasks structure permits dynamic partitioning for a wide class of parallel applications. This paper describes a fault-tolerant implementation of this structure using atomic actions (atomic transactions) to operate on persistent objects, which are accessed in a distributed setting via a Remote Procedure Call (RPC). The system is suited to parallel execution of data and compute intensive programs that require persistent storage and fault tolerance, and runs on stock hardware and software platforms, unix, C++. Its suitability is examined in the context of the measured performance of three applications; ray tracing, matrix multiplication and Cholesky factorization. 1 Introduction Many computations manipulate very large amounts of data. Matrix calculations represent one example class. In a Massively Parallel Processor (MPP) such a vast data set is typically partitioned statically between the very many distributed processing elements and moved amongst them as necessary to perform ...

Citations