A semantically aware transactional concurrency control for GPGPU computing

Abstract

The increased parallel nature of the GPU a ords an opportunity for the exploration of multi-thread computing. With the availability of GPU has recently expanded into the area of general purpose program- ming, a concurrency control is required to exploit parallelism as well as maintaining the correctness of program. Transactional memory, which is a generalised solution for concurrent con ict, meanwhile allow application programmers to develop concurrent code in a more intu- itive manner, is superior to pessimistic concurrency control for general use. The most important component in any transactional memory technique is the policy to solve con icts on shared data, namely the contention management policy. The work presented in this thesis aims to develop a software trans- actional memory model which can solve both concurrent con ict and semantic con ict at the same time for the GPU. The technique di ers from existing CPU approaches on account of the di erent essential ex- ecution paths and hardware basis, plus much more parallel resources are available on the GPU. We demonstrate that both concurrent con-

icts and semantic con icts can be resolved in a particular contention management policy for the GPU, with a di erent application of locks and priorities from the CPU. The basic problem and a software transactional memory solution idea is proposed rst. An implementation is then presented based on the execution mode of this model. After that, we extend this system to re- solve semantic con ict at the same time. Results are provided nally, which compare the performance of our solution with an established GPU software transactional memory and a famous CPU transactional memory, at varying levels of concurrent and semantic con icts.