Research On Last-level Cache Management Strategy For Hybrid Memory System

In recent years,with the rapid development of NVM(Non-Volatile Memory),it has become a trend that DRAM(Dynamic Random Access Memory)and NVM constitute a hybrid storage device to expand the memory of a computer.The traditional cache replacement algorithm does not consider the significant asymmetry between NVM operations and DRAM operations,thus failing to obtain better performance.In response to the above problems,an LLC management solution for hybrid memory architecture is proposed: Missing Aware Partition in Shared LLC(MAPS),which improves system performance while reducing energy consumption.By analyzing the behavior in LLC,it was found that LLC write-backs were unevenly distributed among cache sets.Therefore,a balance mechanism between sets is proposed.Specifically,an access counter and a missing saturation counter are added to each set to distinguish its access and missing characteristics.For the set with frequent access and high missing rate,reduce the overall write-backs of the system by inserting its dirty victim block into the set with infrequent access and low miss rate instead of directly writing back to memory.A type and missing perception of intra-set partition replacement mechanism is proposed.It logically divides each cache set into two dynamically adjusted partitions,each of which caches NVM data and DRAM data.Considering the asymmetrical missing cost of LLC in the hybrid memory system,a new measure is given to calculate the missing rate of each partition weighted according to the number of storage media and occupied blocks,as a basis for dynamically adjusting the size of the partition.Dynamically adjust partition sizes based on workload behavior and balance resource allocation within the set to obtain the best overall performance.A hybrid memory simulation platform was built using gem5 and NVMain,and MAPS was implemented.The SPEC CPU2006 and PARSEC benchmark tests were used to verify the results.The experimental results show that compared with LRU,HAP,and WBAR algorithms,MAPS has increased performance by 44.6%,41.7%,and 34.0%,reducing energy consumption by 15.4%,13.8%,and 10.1%,and extending the life of NVM memory by 229.3%,224.8%,and 195.5%.