Efficient Hot Page Migration Mechanism For Heterogeneous Memory Systems

Traditional memory systems based on Dynamic Random Access Memory(DRAM)have been unable to meet the high bandwidth and large memory requirements of big data applications due to their monotonous structure and limited scalability.A heterogeneous memory system integrating DRAM and high bandwidth memory(HBM)can give full play to the advantages of high bandwidth of HBM and large capacity of DRAM to meet the challenges of limited bandwidth of traditional DRAM memory.Since the data access of applications is often distributed in a power-law manner,in DRAM-HBM parallelized heterogeneous memory,how to reduce the hardware and software overhead of page access monitoring and at the same time achieve efficient data migration is an important problem.Aiming at the problem of page migration in DRAM-HBM parallelized heterogeneous memory,an efficient page migration mechanism based on Majority Element Algorithm(MEA)and random replacement algorithm is proposed,which realizes effective page monitoring through hardware.It mainly completes the following three aspects of work: first,expand the function of the memory controller,integrate the hotspot data monitoring and efficient migration management mechanism in heterogeneous memory,and simply modify its hardware structure design.Complete system performance optimization under the premise of operating system transparency.Secondly,a hotspot data monitoring algorithm based on MEA sorting and a migration strategy based on migration threshold are proposed,which record the historical page access information through a count array of limited size,and use this to efficiently predict its future memory access behavior.The memory access monitoring algorithm is based on the mechanism of page heat decay over time,which can efficiently monitor the dynamic changes of the application working set,identify the hot pages in the DRAM page,and greatly reduce the monitoring overhead of the page.Finally,under the conditions of single-core and multi-core,the performance difference between the random replacement algorithm and other common replacement algorithms is compared and analyzed,and the effectiveness of the optimization mechanism related to hot data monitoring such as the dynamic adjustment strategy of migration threshold in different application scenarios is verified.The most effective hot page migration mechanism in high bandwidth application scenarios.The experimental results show that in the multi-core scenario,compared with the heat monitoring algorithm based on multi-queue,MEA can bring at least 9% improvement effect,and can save 80% of hardware overhead,more effectively monitor and migrate hot pages and reduce hardware design.the complexity.Compared with the traditional LRU replacement algorithm that needs to maintain a large number of pointers,the random replacement algorithm can complete the page replacement within the error range of no more than 0.1% performance loss,and its hardware overhead is almost negligible.