Research On Hybrid--grained Page Management In Hybrid Memory Systems

With the rapid development of computer hardware and software technologies,the traditional Dynamic Random Access Memory(DRAM)has been unable to meet the increasing Memory demand of applications due to its disadvantages,such as large storage energy consumption,small storage density,limited scalability,.Though emerging Non-volatile Memory(NVM)technologies solves these problems perfectly,it cannot be the ideal choice for the memory for its diadvantages such as high access delay,limited write times and high write power consumption.Therefore,hybrid memory systems that combine a small capacity DRAM with large capacity NVM has become a hot topic.However,designing a hybrid memory system with high performance is still facing a dilemma.On one hand,it is needed to support superpages in hybrid memory systems for reducing addressing overhead brings by the large capacity offered by NVMs.On the other hand,it is needed to leverage page migration to improve the system performance of hybrid memory systems.Unfortunately,it is impossible to support efficient page migration and superpages,as coarse-grained page migration brings intolerable overhead while fine-grained page migration undermines address continuity which is necessary for superpages Therefore,in order to design high performance hybrid memory systems,the key challenge is to design a page management mechanism that supporting superpage for low addressing overhead and fine-grained page migration for high performanceTo address these challenges,it is imperative to support multi-grained page management in hybrid memory systems.To achieve this goal,several issues have been addressed.These issues,in particular,mainly include two aspects:the lightweight page monitoring and migration mechanism and the architectural design in hybrid memory systems supporting multi-grained pages management.In terms of lightweight page monitoring and migration mechanism,it is difficult to balance monitoring cost and monitoring accuracy.At the same time,in terms of the architecture design for multi-granularity page migration,the existing mechanism is not good enough to support continuous page mapping,which is embodied in the following aspects:Existing page table management mechanisms are unable to maintain and exploit the inherent continuity of migrated pages;Existing Translation Lookaside Buffer(TLB)mapping mechanisms perform poor performance when mapping these continuous migrated pagesIn order to solve the above research problems and follow the evolution of memory architecture,researches have been carried out in the three following areas:lightweight monitoring and management,page table mechanism optimization and TLB mechanism optimization in systems supporting multi-grained page migrationIn terms of lightweight page monitoring and migration in hybrid memory systems,a novel hybrid memory management mechanism called Rainbow is proposed to support both superpages and lightweight page migration.Rainbow manages the NVM at the superpage granularity,and uses the DRAM to cache frequently-accessed(hot)small pages in the superpages.Correspondingly,Rainbow utilizes split TLBs to support multiple page sizes.A lightweight memory access monitoring mechanism is also proposed to identify the fine-grained hot pages in the NVM,and promote those contiguous,aligned hot pages into a smaller superpage when they are migrated to the DRAM.Rainbow leverages a NVM-to-DRAM address remapping mechanism to identify the migrated small pages,without splintering the superpages.Experimental results show that Rainbow can reduce monitoring costs by up to 98.4%(average 86.3%)without significantly reducing monitoring accuracy.At the same time,Rainbow can significantly reduce the address translation overhead for applications with large memory footprints,and improve application performance by up to 2.9X(45.3%on average)compared to a state-of-the-art memory migration policy without superpage support.In terms of page table design in hybrid memory systems supporting multi-grained page management,a new page-contiguity-aware page Table called SuperPT is proposed to reduce the overheads serving TLB misses.By leveraging a virtual hashed page table and a hybrid DRAM allocator,SuperPT can maintain the page contiguity information before and after page migration,and thus do the address translation in a more flexible and efficient way while rarely destroy the contiguity within migrated pages.Experimental results show that SuperPT can significantly reduce memory access times by 19.3%on average and thus improve the performance by 9.5%on averageIn terms of TLB mechanism optimization in hybrid memory systems supporting multi-grained page management,a new memory management mechanism called Tamp is performed to support multiple page sizes in hybrid memory systems.As there are remarkable levels of intermediate contiguity for hot pages with superpages,a multi-grained TLBs mechanism is proposed to exploit the page contiguity What's more,multi-grained page migration mechanism is also designed to enhance the probability of TLB coalescing.These schemes significantly improve the TLB coverage.Experimental results show that Tamp can significantly reduce the number of TLB misses by 62.4%on average,and thus improve application performance by 16.2%on average,when compared to the other hybrid memory system without multi-grained page management supportIn summary,the monitoring and management mechanism for multi-grained page migration,the optimization mechanism of page table for multi-grained page migration,and the optimization mechanism of TLB architecture for multi-granularity page migration are proposed in hybrid memory systems,which solve the shortcomings of existing researches from different aspects.