Virtual Memory Paging Mechanism Based On Distributed Heterogeneous Memory Pool

In the cloud environment,disaggregated memory decouples computing resources and memory resources in a cluster by accessing remote memory.This improves memory utilization while improving task throughput.However,in a disaggregated memory system,the remote memory access is affected by the transmission overhead on the critical path,and the latency is generally high.Moreover,the data access method based on paging often brings the problem of read and write amplification,which further increases the remote memory access latency.It is of great significance to study the new disaggregated memory access mechanism to reduce the data remote access latency and improve the performance of big data applications.VANDI,a virtual memory paging mechanism based on heterogeneous memory pool,was proposed to solve the problem of high latency of disaggregated memory,and optimized from two aspects of cache and prefetch.In terms of caching mechanism,firstly,write operations are asynchronized through multiple buffer queues to reduce the access latency on the critical path.Secondly,two logical cache queues with different capacity are constructed in the single-layer cache.Data status is updated according to the time and space interval of data access,and migration is carried out in the logical queue at the same time,which can effectively retain hot data and improve cache hit ratio.In terms of data prefetching,first of all,according to the read access history of the swap partition,it is judged whether the addresses are continuous,and whether there is aggregation in the case of nonstrict continuous.In this way,the possible sequential access trend is predicted and data in the remote memory pool is prefetched synchronously.Second,prefetch trigger conditions are dynamically changed based on cache hit ratio and prefetch coverage changes.At the same time,the prefetch granularity is adaptively changed according to the data matching situation,and the prefetch data and access data are managed hierarchically.Different cache replacement strategies are adopted to improve cache hit ratio and reduce cache pollution caused by prefetching.The experiment tests the overall performance of the system and the effectiveness of the core strategy on the real host computers.Compared to Infiniswap,VANDI delivers 6.8 times better performance in sequential access mode and 15 times better performance in random access mode.VANDI still has a performance improvement of 1.2 to 1.5 times over Valet,a remote memory system with caching.In tests of applications,VANDI showed performance improvements of 20% to 80% over Valet.Validity tests also verify that caching and prefetching mechanisms improve the cache hit ratio,reaching 95% at the zipfian parameter of 1.2.