A Taxonomy and Design Methodology for Hybrid Memory Systems

Licentiate thesis, 2014

The number of concurrently executing processes and their memory demand in multicore systems continue to grow. Larger and still fast main memory is needed for meeting the demand and avoiding an increase in backing store accesses that are much slower and less energy efficient than main memory accesses. Luckily, Non-Volatile Memory (NVM) technologies can bridge the cost, density, performance, and energy efficiency gaps between backing store and DRAM, the conventional main memory technology. Thus, NVM can be combined with DRAM into hybrid main memory striving to enjoy both the larger capacity enabled by NVM and the speed and energy efficiency of DRAM. NVM adds a new dimension to the system design space inspiring researchers to investigate sophisticated hybrid memories. This has resulted in a large body of work that, unfortunately, lacks systematization. The thesis at hand addresses this problem by proposing a taxonomy and a notation for classifying hybrid main memory organizations. The design space of hybrid systems is large, and the best partitioning of resources between DRAM and NVM is nontrivial. The high implementation and computation efforts of detailed modeling impede extensive design space exploration required for finding the most promising design points. This thesis aids such extensive exploration by proposing a workload methodology and first-order models for system-level execution time and energy. Next, the thesis contributes with Rock, an insightful performance model showing how memory system throughput can be boosted by installing more DRAM and NVM thus motivating Design-time Resource Partitioning (DRP). The lack of an approach suitable for extensive partitioning is addressed by proposing Crystal, a DRP method powered by the system-level models and framing partitioning as an optimization problem, such that the first-order nature of the models does not restrict its applicability, as shown by validation. Crystal is practical and facilitates early and rapid DRP finding promising design points for further detailed evaluation. For instance, Crystal shows how for specific workloads higher performance and energy efficiency can be achieved by employing NVM with the speed and energy consumption of NAND Flash instead of a much faster and more energy efficient NVM technology like phase-change memory.