A Fault – Tolerant System for Balancing the Load of Data – Parallel Applications

Abstract –– In distributed computing environments, fault–tolerance is an important objective, especially for parallel applications. Many distributed computing environments achieve fault–tolerance by periodic checkpointing. This has the advantage of relative ease of implementation and can be considered equivalent to task migration. However, there are two main disadvantages of such environments. One is that any work in progress after checkpointing is lost when a fault occurs. The other is that these systems are heavily reliant on task migration as the only mechanism for load balancing. This paper presents a system that overcomes these shortcomings by task duplication and by the integration of data migration into task migration as a load balancing mechanism. It also presents results of a preliminary implementation.