Design and Implementation of an Efficient and Scalable Software Distributed Shared Memory System
This thesis presents the design and implementation of our novel hybrid software DSM system. We call our system hybrid home-based EAC (HHEAC) since the system implements our novel exclusive access consistency model (EAC) based on the hybrid protocol of the homeless and home-based protocols. HHEAC guarantees only that shared variables inside a critical section are up to date before the accesses. Other shared variables outside a critical section are guaranteed to be up to date after the next barrier synchronisation. Our home-based DSM implementation is different from the previous implementations in that a home node does not receive any diffs from non-home nodes until the next barrier synchronisation. It is also different in that during a lock synchronisation required diffs are prefetched before the critical section, which reduces not only data traffic but also page faults inside the critical section. We also present a diff integration technique that can further unnecessary data traffic during lock synchronisation. This technique is especially effective in reducing data traffic for migratory applications. Finally, we develop a home migration technique that solves the wrong home assignment problem in the home-based protocol. Our technique is different from others in that an optimum home node is decided before updating a home node. To evaluate our system, we performed various experiments with well-known benchmark applications, including a novel parallel neural network application.The performance evaluation shows that HHEAC is more scalable than other DSM systems such as TreadMarks and removes the home assignment problem in the conventional home-based protocol.
Advisor: Cranefield, Stephen; Purvis, Martin; Werstein, Paul
Degree Name: Doctor of Philosophy
Degree Discipline: Computer Science
Keywords: Distributed Shared Memory; High Performance Computing; Parallel Computing; Cache Coherence
Research Type: Thesis