Shared Resource Management for Non-Volatile Asymmetric Memory

摘要

Non-volatile memory (NVM), such as Phase-Change Memory (PCM), is a promising energy-efficient candidate to replace DRAM. It is desirable because of its non-volatility, good scalability and low idle power. NVM, nevertheless, faces important challenges. The main problems are: writes are much slower and more power hungry than reads and write bandwidth is much lower than read bandwidth. Hybrid main memory architecture, which consists of a large NVM and a small DRAM, may become a solution for architecting NVM as main memory. Adding an extra layer of cache mitigates the drawbacks of NVM writes. However, writebacks from the last-level cache (LLC) might still (a) overwhelm the limited NVM write bandwidth and stall the application, (b) shorten lifetime and (c) increase energy consumption.ud udEffectively utilizing shared resources, such as the last-level cache and the memory bandwidth, is crucial to achieving high performance for multi-core systems. No existing cache and bandwidth allocation scheme exploits the read/write asymmetry property, which is fundamental in NVM. This thesis tries to consider the asymmetry property in partitioning the cache and memory bandwidth for NVM systems.ud udThe thesis proposes three writeback-aware schemes to manage the resources in NVM systems. First, a runtime mechanism, Writeback-aware Cache Partitioning (WCP), is proposed to partition the shared LLC among multiple applications. Unlike past partitioning schemes, WCP considers the reduction in cache misses as well as writebacks. Second, a new runtime mechanism, Writeback-aware Bandwidth Partitioning (WBP), partitions NVM service cycles among applications. WBP uses a bandwidth partitioning weight to reflect the importance of writebacks (in addition to LLC misses) to bandwidth allocation. A companion Dynamic Weight Adjustment scheme dynamically selects the cache partitioning weight to maximize system performance. Third, Unified Writeback-aware Partitioning (UWP) partitions the last-level cache and the memory bandwidth cooperatively. UWP can further improve the system performance by considering the interaction of cache partitioning and bandwidth partitioning. The three proposed schemes improve system performance by considering the unique read/write asymmetry property of NVM.