CoD: Coherence-on-Demand — Runtime Adaptable Working Set Coherence for DSM-Based Manycore Architectures

摘要

Embedded system applications, with their inherently limited parallelism, rarely exploit all available processing resources in large DSM-based manycore architectures. In addition, global coherence spanning across all tiles does not scale well. Therefore, we have proposed a region-based cache coherence (RBCC) approach that enables coherence among a selectable cluster of tiles in accordance with application requirements. In this paper, we present a novel RBCC-malloc() extension that transparently tailors coherence to actually shared application working sets at runtime. Further, the design and hardware implementation of a flexibly configurable coherency region manager (CRM) supporting RBCC-malloc() are introduced. We synthesized the CRM on an FPGA for a 64-core system and observed a 57% reduction in BRAM-utilization compared to a global coherence directory for regions with up to 16 cores. Experiments reveal an application acceleration of up to 42% compared to a message passing based implementation. We also demonstrate the advantage of RBCC-malloc() compared to standalone RBCC.