Efficient Data Migration For Heterogeneous Memory System

Nowadays,heterogeneous memory that consists of DRAM and NVM becomes more and more popular.To satisfy different applications’ demands,the unified management of data migration between DRAM and NVM is necessary,including data placement and swap in the parallel architecture,and data replacement,eviction,and copy in the hierarchical architecture.At present,there are three typical data migration mechanisms in different scenarios.First,cache data migration is usually used for high-performance computing.NVM is used as the main memory and DRAM is used as the cache of NVM.It aims to get high performance in various ways,e.g,improving the cache hit ratio.Influenced by inaccurate access predictions and coarse grain management,cache data migration faces the huge migration overhead and low cache space utilization,limiting the further improvement of performance.Second,checkpointing data migration is usually used for persistent memory.It migrates data from volatile memory to NVM and builds memory checkpoints periodically for backup.It has two performance bottlenecks in the critical path:the large amount of NVM writes and metadata overhead.Limited by the small hardware space,the two bottlenecks cannot be solved at the same time,introducing a difficult tradeoff.Third,dirty data migration is usually used for big data storage.DRAM is used as the cache to filter the write traffic of NVM to prolong its lifetime.Influenced by space competitions and inaccurate dirty data predictions,different optimization policies conflict frequently,introducing significant negative effects on performance and NVM lifetime.Data migration plays a key role in heterogeneous memory management to satisfy the demands in different scenarios.This dissertation presents research on how to design and implement efficient data migration for heterogeneous memory systems.The main contributions are shown below:●In the cache data migration,Object Fingerprint Cache is proposed to improve DRAM cache utilization and reduce memory migration overhead.It facilitates accurate migrations of the object’s CPU cachelines from NVM to DRAM cache.To make the object migration more practical,an online objects memory access behavior collecting method is proposed.To improve DRAM cache space utilization,dense cache sectors with multiple granularities are devised,which organize sectors with different cache partitions,according to their sizes.In addition,the capacities of cache partitions are able to be dynamically adjusted to adapt to the workload phase changes.Compared with state-of-the-art hardware caches,the object fingerprint cache gets 48.7%migration overhead reduction and 22.7%IPC improvement.●In the checkpointing data migration,an architectural approach,Dual-Page Checkpointing is proposed to support memory data persistence on NVM.Dual-page mapping maintains address mappings at both page granularity and cacheline granularity in an integrated manner without incurring excessive space and management overhead.It solves the traditional dilemma in the trade-off between metadata hardware overhead and extra NVM data writes.It demonstrates that a simple hardware-based checkpointing mechanism in the memory controller can efficiently achieve data durability and crash consistency of in-memory data and avoid the overhead of complex state-of-the-art software solutions.The evaluation shows that dual-page checkpointing outperforms the lattest software solution by 13.6x higher throughput.Compared with the state-of-the-art checkpointing techniques,it gets 89.5%metadata overhead reduction,33.7%NVM writes reduction,and 27.7%performance improvement.●In the dirty data migration,Victim-History-Aware Cache is proposed to coordinate the performance and lifetime optimization policies to reduce conflicts.Based on the eviction patterns of DRAM cache pages,a behavior prediction method is proposed to estimate the cost-benefit of data migration accurately.To reduce space competitions,the management can be dynamically adjusted to adapt to the various request behaviors.To reduce the conflicts between optimization policies,a victimhistory-aware replacement algorithm is devised to choose the best victim page by considering multiple factors comprehensively.Compared with state-of-the-art systems,it gets 10.5%NVM lifetime and 12.9%performance improvement.