An approach to fault-tolerant parallel processing on intermittently idle, heterogeneous workstations

Abstract

W e propose a novel approach to harness the idle cycles of workstations connected by L A N / W A N s for long running scientific computations and present performance results for our prototype system called Persistent Linda (PLinda). PLinda offers low runtime overhead and, migration among heterogeneous architectures while retreating quickly when owners return to their workstations. PLinda achieves this by implementing a lightweight transaction model that lacks serializability and durability, but preserves a basic guarantee: if a PLinda execution terminates, it has the same result as some failure-free Linda execution. Further, by storing the state of a PLinda process as a set of core variables as of each transaction commit, a PLinda process can migrate among different architectures.Within the space of lightweight transaction models, we offer three mechanisms that make different tradeoffs between failure-free performance and recovery time. All three mechanisms may be used at the same time in a single application; each process using the mechanism which is best suited for its characteristics. Our experiments illustrate the tradeoffs of the three mechanisms as well as the overall performance of the system on applications from physics and finance.