Memory Management for Emerging Memory Technologies
MetadataShow full item record
The Memory Wall, or the gap between CPU speed and main memory latency, is ever increasing. The latency of Dynamic Random-Access Memory (DRAM) is now of the order of hundreds of CPU cycles. Additionally, the DRAM main memory is experiencing power, performance and capacity constraints that limit process technology scaling. On the other hand, the workloads running on such systems are themselves changing due to virtualization and cloud computing demanding more performance of the data centers. Not only do these workloads have larger working set sizes, but they are also changing the way memory gets used, resulting in higher sharing and increased bandwidth demands. New Non-Volatile Memory technologies (NVM) are emerging as an answer to the current main memory issues. This thesis looks at memory management issues as the emerging memory technologies get integrated into the memory hierarchy. We consider the problems at various levels in the memory hierarchy, including sharing of CPU LLC, traffic management to future non-volatile memories behind the LLC, and extending main memory through the employment of NVM. The first solution we propose is “Adaptive Replacement and Insertion" (ARI), an adaptive approach to last-level CPU cache management, optimizing the cache miss rate and writeback rate simultaneously. Our specific focus is to reduce writebacks as much as possible while maintaining or improving miss rate relative to conventional LRU replacement policy, with minimal hardware overhead. ARI reduces writebacks on benchmarks from SPEC2006 suite on average by 32.9% while also decreasing misses on average by 4.7%. In a PCM based memory system, this decreases energy consumption by 23% compared to LRU and provides a 49% lifetime improvement beyond what is possible with randomized wear-leveling. Our second proposal is “Variable-Timeslice Thread Scheduling" (VATS), an OS kernel-level approach to CPU cache sharing. With modern, large, last-level caches (LLC), the time to fill the LLC is greater than the OS scheduling window. As a result, when a thread aggressively thrashes the LLC by replacing much of the data in it, another thread may not be able to recover its working set before being rescheduled. We isolate the threads in time by increasing their allotted time quanta, and allowing larger periods of time between interfering threads. Our approach, compared to conventional scheduling, mitigates up to 100% of the performance loss caused by CPU LLC interference. The system throughput is boosted by up to 15%. As an unconventional approach to utilizing emerging memory technologies, we present a Ternary Content-Addressable Memory (TCAM) design with Flash transistors. TCAM is successfully used in network routing but can also be utilized in the OS Virtual Memory applications. Based on our layout and circuit simulation experiments, we conclude that our FTCAM block achieves an area improvement of 7.9× and a power improvement of 1.64× compared to a CMOS approach. In order to lower the cost of Main Memory in systems with huge memory demand, it is becoming practical to extend the DRAM in the system with the less-expensive NVMe Flash, for a much lower system cost. However, given the relatively high Flash devices access latency, naively using them as main memory leads to serious performance degradation. We propose OSVPP, a software-only, OS swap-based page prefetching scheme for managing such hybrid DRAM + NVM systems. We show that it is possible to gain about 50% of the lost performance due to swapping into the NVM and thus enable the utilization of such hybrid systems for memory-hungry applications, lowering the memory cost while keeping the performance comparable to the DRAM-only system.
Fedorov, Viacheslav (2016). Memory Management for Emerging Memory Technologies. Doctoral dissertation, Texas A&M University. Available electronically from