A Taxonomy and Design Methodology for Hybrid Memory Systems
Dmitry Knyaginin · Chalmers Research (Chalmers University of Technology) · 2014
The number of concurrently executing processes and their memory demandin multicore systems continue to grow. Larger and still fast mainmemory is needed for meeting the demand and avoiding an increase inbacking store accesses that are much slower and less energy efficientthan main memory accesses. Luckily, Non-Volatile Memory (NVM)technologies can bridge the cost, density, performance, and energyefficiency gaps between backing store and DRAM, the conventional mainmemory technology. Thus, NVM can be combined with DRAM into hybridmain memory striving to enjoy both the larger capacity enabled by NVMand the speed and energy efficiency of DRAM. NVM adds a new dimensionto the system design space inspiring researchers to investigatesophisticated hybrid memories. This has resulted in a large body ofwork that, unfortunately, lacks systematization. The thesis at handaddresses this problem by proposing a taxonomy and a notation forclassifying hybrid main memory organizations. The design space ofhybrid systems is large, and the best partitioning of resourcesbetween DRAM and NVM is nontrivial. The high implementation andcomputation efforts of detailed modeling impede extensive design spaceexploration required for finding the most promising designpoints. This thesis aids such extensive exploration by proposing aworkload methodology and first-order models for system-level executiontime and energy. Next, the thesis contributes with Rock, an insightfulperformance model showing how memory system throughput can be boostedby installing more DRAM and NVM thus motivating Design-time ResourcePartitioning (DRP). The lack of an approach suitable for extensivepartitioning is addressed by proposing Crystal, a DRP method poweredby the system-level models and framing partitioning as an optimizationproblem, such that the first-order nature of the models does notrestrict its applicability, as shown by validation. Crystal ispractical and facilitates early and rapid DRP finding promising designpoints for further detailed evaluation. For instance, Crystal showshow for specific workloads higher performance and energy efficiencycan be achieved by employing NVM with the speed and energy consumptionof NAND Flash instead of a much faster and more energy efficient NVMtechnology like phase-change memory.