Evolve : a preliminary multicore architecture for Introspective Computing

Abstract

This thesis creates a framework for Introspective Computing. Introspective Computing is a computing paradigm characterized by self-aware software. Self-aware software systems use hardware mechanisms to observe an application's execution so that they may adapt execution to improve performance, reduce power consumption, or balance user-defined fitness criteria over time-varying conditions in a system environment. We dub our framework Partner Cores. The Partner Cores framework builds upon tiled multicore architectures [11, 10, 25, 9], closely coupling cores such that one may be used to observe and optimize execution in another. Partner cores incrementally collect and analyze execution traces from code cores then exploit knowledge of the hardware to optimize execution. This thesis develops a tiled architecture for the Partner Cores framework that we dub Evolve. Evolve provides a versatile substrate upon which software may coordinate core partnerships and various forms of parallelism. To do so, Evolve augments a basic tiled architecture with introspection hardware and programmable functional units. Partner Cores software systems on the Evolve hardware may follow the style of helper threading [13, 12, 6] or utilize the programmable functional units in each core to evolve application-specific coprocessor engines. This thesis work develops two Partner Cores software systems: the Dynamic Partner-Assisted Branch Predictor and the Introspective L2 Memory System (IL2). The branch predictor employs a partner core as a coprocessor engine for general dynamic branch prediction in a corresponding code core. The IL2 retasks the silicon resources of partner cores as banks of an on-chip, distributed, software L2 cache for code cores.

(cont.) The IL2 employs aggressive, application-specific prefetchers for minimizing cache miss penalties and DRAM power consumption. Our results and future work show that the branch predictor is able to sustain prediction for code core branch frequencies as high as one every 7 instructions with no degradation in accuracy; updated prediction directions are available in a low minimum of 20-21 instructions. For the IL2, we develop a pixel block prefetcher for the image data structure used in a JPEG encoder benchmark and show that a 50% improvement in absolute performance is attainable.