Research And Optimization Of Reading And Writing Strategies For Software Transactional Memory System Based On RSTM

In the future,the parallel development remains a major challenge faced by programmers.ILP limitations and capacity constraints are forcing manufacturers to study the multicore chip instead of enhancing extending a single processor core performance.Chip Multi-processors(CMPs)is used in many applications.There is a series of problems in the traditional lock mechanism.Under this situation,transactional memory comes into being.Transactional memory allows that atomic transaction blocks to execute concurrently,while execute serially when transactions conflict happens.Transactional memory system has a new architecture for multi-core and provides an efficient parallel programming environment for programmers.Using transaction can solve a series of problems which are caused by lock mechanism,and improve its concurrency.In this thesis,the status of transactional memory is studied,and its reading and writing behaviors are studied deeply.Traditional reading and writing strategies include reading validation and writing acquisition.These strategies have a very important influence on the performance of the system.The current traditional reading validation and writing acquisition strategies are analyzed in the thesis,but none of them can fit all the applications alone.Therefore,the strategy is improved according to the data structure of RSTM and the status of running processes.A dynamic adaptive reading validation algorithm called DARV is designed.It can adjust the validation strategies referring to the performance of the application and the number of the transactions dynamically.Some benchmarks of RSTM are used to test the performance of the dynamic adaptive read validation strategy DARV.Experimental results show that DARV has a lower transaction abort rather than the traditional reading validations.Writing acquisition is another important factor to impact transaction processing.The writing acquisition strategies in RSTM are studied,and a dynamic adaptive lock acquisition strategy called DALA is proposed and implemented.It can choose different writing acquisitions for different cases by the performance of the transaction and the nmber of the transactions.Some benchmarks of RSTM are used to test the performance of DALA.Experimental results show that DALA has better performance in most cases.In this thesis,the research background and related technologies are introduced at first.Then two strategies about transactional reading and writing are proposed.They are dynamic adaptive read validation(DARV)and dynamic adaptive lock acquisition(DALA).The design and implementation are described in detail.And their performances are tested.Finally,our work is summarized and the work in future is prospected.