Chen, Xi (2013-05). Physical Planning and Uncore Power Management for Multi-Core Processors. Doctoral Dissertation.
For the microprocessor technology of today and the foreseeable future, multi-core is a key engine that drives performance growth under very tight power dissipation constraints. While previous research has been mostly focused on individual processor cores, there is a compelling need for studying how to efficiently manage shared resources among cores, including physical space, on-chip communication and on-chip storage.
In managing physical space, floorplanning is the first and most critical step that largely affects communication efficiency and cost-effectiveness of chip designs. We consider floorplanning with regularity constraints that requires identical processing/memory cores to form an array. Such regularity can greatly facilitate design modularity and therefore shorten design turn-around time. Very little attention has been paid to automatic floorplanning considering regularity constraints because manual floorplanning has difficulty handling the complexity as chip core count increases. In this dissertation work, we investigate the regularity constraints in a simulated-annealing based floorplanner for multi/many core processor designs. A simple and effective technique is proposed to encode the regularity constraints in sequence-pair, which is a classic format of data representation in automatic floorplanning. To the best of our knowledge, this is the first work on regularity-constrained floorplanning in the context of multi/many core processor designs.
On-chip communication and shared last level cache (LLC) play a role that is at least as equally important as processor cores in terms of chip performance and power. This dissertation research studies dynamic voltage and frequency scaling for on-chip network and LLC, which forms a single uncore domain of voltage and frequency. This is in contrast to most previous works where the network and LLC are partitioned and associated with processor cores based on physical proximity. The single shared domain can largely avoid the interfacing overhead across domain boundaries and is practical and very useful for industrial products. Our goal is to minimize uncore energy dissipation with little, e.g., 5% or less, performance degradation. The first part of this study is to identify a metric that can reflect the chip performance determined by uncore voltage/frequency. The second part is about how to monitor this metric with low overhead and high fidelity. The last part is the control policy that decides uncore voltage/frequency based on monitoring results. Our approach is validated through full system simulations on public architecture benchmarks.