Addressing Prolonged Restore Challenges in Further Scaling Drams

摘要

As the de facto memory technology, DRAM has enjoyed continuous scaling over the past decades to keep performance growth and capacity enhancement. However, DRAM further scaling into deep sub-micron regime faces significant challenges. Among the induced issues, prolonged restore time is expected to be one of the major concerns, but it has been paid little attention. Aiming at restore issue, this thesis performs pioneering studies to characterize the problems, and presents techniques from different perspectives to overcome them.;First, our experimental studies quantify the significant restore process variations, caus- ing serious degradations on yield and/or performance. To solve the problem, we propose schemes to expose the variations to the architectural levels. Fast restore chunks can thus be constructed utilizing DRAM organization, and they can be exposed to the memory controller to effectively compensate the performance loss. Further, we maximize the improvement by applying restore-time-aware rank construction and hotness-aware page allocation schemes to fully utilize the fast regions.;Second, in addition to simply expose the variations to higher levels, we investigate DRAM cell structures and behaviors finding that refresh and restore are two strongly correlated operations. Whereas are being fully restored after each read or write access, DRAM cells are always being fully charged by periodical refresh operations, providing an opportunity to early terminate restore. With the insight, we first propose to truncate a restore using the time distance to next refresh. Further, to provide more truncation opportunities, we integrate the multirate-refresh concepts to shorten the distance by increasing the refresh rate of recently accessed regions.;Lastly, we explore higher to the application level with the inspiration that a large set of applications can well tolerate output accuracy loss and runtime errors, enabling us to exploit approximate computing to mitigate prolonged restore. By utilizing the variance in restore timing exhibited at different row segments, we reduce the restore time such that only partial segments are fully reliable. We then map the critical data onto the reliable segments to keep the application-level errors low. Atop of the approximation-aware technique, we further generalize it to support precise computing as well.