Integration of Non-volatile Memory into Storage Hierarchy
MetadataShow full item record
In this dissertation, we present novel approaches for integrating non-volatile memory devices into storage hierarchy of a computer system. There are several types of non- volatile memory devices, such as flash memory, Phase Change Memory (PCM), Spin- transfer torque memory (STT-RAM). These devices have many appealing features for applications; however, they also offer several challenges. This dissertation is focused on how to efficiently integrate these non-volatile memories into existing memory and disk storage systems. This work is composed of two major parts. The first part investigates a main-memory system employing Phase Change Memory instead of traditional DRAM. Compared to DRAM, PCM has higher density and no static power consumption, which are very important factors for building large capacity memory systems. However, PCM has higher write latency and power consumption compared to read operations. Moreover, PCM has limited write endurance. To efficiently integrate PCM into a memory system, we have to solve the challenges brought by its expensive write operations. We propose new replacement policies and cache organizations for the last-level CPU cache, which can effectively reduce the write traffic to the PCM main memory. We evaluated our design with multiple workloads and configurations. The results show that the proposed approaches improve the lifetime and energy consumption of PCM significantly. The second part of the dissertation considers the design of a data/disk storage using non-volatile memories, e.g. flash memory, PCM and nonvolatile DIMMs. We consider multiple design options for utilizing the nonvolatile memories in the storage hierarchy. First, we consider a system that employs nonvolatile memories such as PCM or nonvolatile DIMMs on memory bus along with flash-based SSDs. We propose a hybrid file system, NVMFS, that manages both these devices. NVMFS exploits the nonvolatile memory to improve the characteristics of the write workload at the SSD. We satisfy most small random write requests on the fast nonvolatile DIMM and only do large and optimized writes on SSD. We also group data of similar update patterns together before writing to flash-SSD; as a result, we can effectively reduce the garbage collection overhead. We implemented a prototype of NVMFS in Linux and evaluated its performance through multiple benchmarks. Secondly, we consider the problem of using flash memory as a cache for a disk drive based storage system. Since SSDs are expensive, a few SSDs are designed to serve as a cache for a large number of disk drives. SSD cache space can be used for both read and write requests. In our design, we managed multiple flash-SSD devices directly at the cache layer without the help of RAID software. To ensure data reliability and cache space efficiency, we only duplicated dirty data on flash- SSDs. We also balanced the write endurance of different flash-SSDs. As a result, no single SSD will fail much earlier than the others. Thirdly, when using PCM-like devices only as data storage, it’s possible to exploit memory management hardware resources to improve file system performance. However, in this case, PCM may share critical system resources such as the TLB, page table with DRAM which can potentially impact PCM’s performance. To solve this problem, we proposed to employ superpages to reduce the pressure on memory management resources. As a result, the file system performance is further improved.
Qiu, Sheng (2013). Integration of Non-volatile Memory into Storage Hierarchy. Doctoral dissertation, Texas A & M University. Available electronically from