Abstract
Owing to the advantages of low standby power and high scalability, ReRAM technology is considered as a promising replacement for conventional DRAM in future manycore systems. In order to make ReRAM highly scalable, the memory array has to have a crossbar array structure, which needs a specific access mechanism for activating a row of memory when reading/writing a data block from/to it. This type of memory access would cause Sneak Current that would lead to voltage drop on the memory cells of the activated row, i.e., the cells which are far from the write drivers experience more voltage drop compared to those close to them. This results in a nonuniform access latency for the cells of the same row. To address this problem, we propose and evaluate a scheme that exploits the non-uniformity of write access pattern of the workloads. More specifically, based on our extensive characterization of write patterns to the cache lines and memory pages of 20 CPU workloads, we recognized that (i) on each main memory access, just a few cache lines of the activated row need to be updated on a write-back, and more importantly, there is a temporal and spatial locality of the writes to the cache lines; and (ii) all pages of the memory footprint of an application do not see the same write counts during the execution of the workload. Motivated by these characteristics, we then evaluate different intra-page memory block permutations in order to improve the performance of a crossbar ReRAM-based main memory. Our results collectively show that, by applying some types of intra-page memory block permutation, the access latency to a ReRAM-based main memory can be reduced up to 50% when running the SPEC CPU2006 workloads.