Research On Scalability Of Durable Software Transactional Memory Based On Timestamp

The emerging non-volatile memory(NVM)provides a lot of advantages,including byte-addressability,durability,large capacity and low energy consumption.Based on these features,NVM is expected to be widely used in the future.However,programming on NVM is very complex,especially in the current multi-core environment,because users should not only ensure the consistency of data crash,but also ensure the correctness of concurrency.In order to solve this problem,combining persistence and traditional transactional memory,researchers have proposed a persistent transactional memory solution.According to different implementation methods,existing persistent transactional memory solutions can be divided into two categories: hardware-assisted persistent transactional memory and persistent software transactional memory.The persistent software transaction memory does not depend on specific hardware,so it has good compatibility and portability,but its scalability is subject to certain restrictions.In response to this problem,the factors affecting the memory scalability of persistent software transactions were analyzed through testing,and it was found that the global logical clock and redundant NVM write operations severely restricted its scalability.In order to eliminate the impact of these two factors on scalability: a thread logical clock method is proposed,which eliminates the problem of global logical clock centralization by allowing each thread to own its independent clock;a cacheline-concious dual version method is proposed,and when updating data,the older version is directly overwritten while the newer version is used to restore data in the event of a crash,thereby eliminating redundant NVM write operations.And based on these two methods,a time-stamp-based durable software transactional memory(Scalable Durable Transactional Memory,SDTM)was implemented.Based on the validity test of the thread logical clock method and the cacheline-concious dual version method,three different concurrent and persistent data structures are implemented based on SDTM,and the load of different read and write ratios is used for testing.It is compared with DudeTM and PMDK.In comparison,its throughput has been improved by 39 times and 146 times respectively.In addition,the real load generated by YCSB is used to test the B+ tree,which is based on SDTM.The test results show that compared with DudeTM and PMDK,its throughput is increased by up to 2.8 times and 29 times,respectively.