Research On Transaction Optimization Based On Persistent Memory Dynamic Library

In recent years,the new Non-Volatile Memory(NVM)devices was born because of the development of hardware technology.Compared with traditional mechanical hard disk(HDD),NVM devices combines the byte-addressability with the persistence,but they also have some disadvantages,such as DRAM-comparable access latency but high write latency,low durability.In order to provide fully utilization to NVM devices,researchers both in industry and academia developed a number of new programming models,which mainly include three ways: the research of native persistence index structure,a new type of persistent memory file system and new persistent memory programming model.Persistent Memory Development Kit(PMDK)is The most representative library programming model.In this paper,our research focus on the transaction mechanism in PMDK,and make the following contributions:(1)The existing transaction mechanism of PMDK lead to poor performance.To figure this reason out,we designed some targeted tests with frequent operations,based on the study we analyzed the process of each operation's execution,and we founded that the performance degradation is attributed to two reasons.On the one hand,persistent memory deallocation incurs high impact on the overall transaction performance,since it makes a redo logging to guarantee the atomicity of persistent memory deallocation.On the other hand,it will try to reuse the current active memory block in thread heap immediately once the free space is exhausted,and memory allocation requests will be suspended during memory recycling.Due to such issues,it makes the flexibility of persistent memory recycling reduce,make the efficiency of free space management reduce,and lead to poor performance of persistent memory allocation.(2)The performance of PMDK comes to a bottleneck caused by memory deallocation.To speedup persistent memory management,we propose DeallocationFree Transaction(DFTX),a transaction mechanism without deallocation.It reduces the overhead of transaction performance on critical path by removing persistent memory deallocate operation from transactions in foreground threads.Experimental results show that DFTX improves the performance of deallocation by 42.03% to 57.52% in single-thread model and 44.54% to 59.05% in multi-thread model compared with the way in PMDK.And DFTX also make improvements to the write amplification on NVM when data updating occurs on fined-grained memory,it can reduce the write amplification by 99.9% and it improves system performance by 20%.(3)The flexibility of memory recycle mechanism in PMDK is insufficient.To address such issue,we propose an optimized mechanism with dynamic approach to recycling memory based on DFTX.It registers information of deallocated memory blocks in a hash table,until it can be recycled.Experimental results show that the optimized mechanism improves the performance of memory recycle compared with the way in PMDK,it improves the system throughput by 11.12% to 97.88% in different thread models.