Embedded dynamic programming networks for networks-on-chip

Abstract

Relentless technology downscaling and recent technological advancements in three dimensional integrated circuit (3D-IC) provide a promising prospect to realize heterogeneous system-on-chip (SoC) and homogeneous chip multiprocessor (CMP) based on the networks-onchip (NoCs) paradigm with augmented scalability, modularity and performance. In many cases in such systems, scheduling and managing communication resources are the major design and implementation challenges instead of the computing resources. Past research efforts were mainly focused on complex design-time or simple heuristic run-time approaches to deal with the on-chip network resource management with only local or partial information about the network. This could yield poor communication resource utilizations and amortize the benefits of the emerging technologies and design methods. Thus, the provision for efficient run-time resource management in large-scale on-chip systems becomes critical. This thesis proposes a design methodology for a novel run-time resource management infrastructure that can be realized efficiently using a distributed architecture, which closely couples with the distributed NoC infrastructure. The proposed infrastructure exploits the global information and status of the network to optimize and manage the on-chip communication resources at run-time. There are four major contributions in this thesis. First, it presents a novel deadlock detection method that utilizes run-time transitive closure (TC) computation to discover the existence of deadlock-equivalence sets, which imply loops of requests in NoCs. This detection scheme, TC-network, guarantees the discovery of all true-deadlocks without false alarms in contrast to state-of-the-art approximation and heuristic approaches. Second, it investigates the advantages of implementing future on-chip systems using three dimensional (3D) integration and presents the design, fabrication and testing results of a TC-network implemented in a fully stacked three-layer 3D architecture using a through-silicon via (TSV) complementary metal-oxide semiconductor (CMOS) technology. Testing results demonstrate the effectiveness of such a TC-network for deadlock detection with minimal computational delay in a large-scale network. Third, it introduces an adaptive strategy to effectively diffuse heat throughout the three dimensional network-on-chip (3D-NoC) geometry. This strategy employs a dynamic programming technique to select and optimize the direction of data manoeuvre in NoC. It leads to a tool, which is based on the accurate HotSpot thermal model and SystemC cycle accurate model, to simulate the thermal system and evaluate the proposed approach. Fourth, it presents a new dynamic programming-based run-time thermal management (DPRTM) system, including reactive and proactive schemes, to effectively diffuse heat throughout NoC-based CMPs by routing packets through the coolest paths, when the temperature does not exceed chip’s thermal limit. When the thermal limit is exceeded, throttling is employed to mitigate heat in the chip and DPRTM changes its course to avoid throttled paths and to minimize the impact of throttling on chip performance. This thesis enables a new avenue to explore a novel run-time resource management infrastructure for NoCs, in which new methodologies and concepts are proposed to enhance the on-chip networks for future large-scale 3D integration.