טכניון מכון טכנולוגי לישראל
הטכניון מכון טכנולוגי לישראל - בית הספר ללימודי מוסמכים  
Ph.D Thesis
Ph.D StudentManevich Ran
SubjectCentralized Paradigms in Network on Chip Architectures
DepartmentDepartment of Electrical Engineering
Supervisors Professor Emeritus Israel Cidon
Professor Emeritus Avinoam Kolodny
Full Thesis textFull thesis text - English Version


Abstract

In our research, we evaluate the feasibility and the potential benefits of the utilization of centralized paradigms in Networks-on-Chip (NoCs). We propose a bus-NoC hybrid architecture termed BENoC (Bus Enhanced Network-on-Chip) that can be used as a platform for centralized off-band NoC control mechanisms. We introduce ATDOR (Adaptive Toggle Dimension Ordered Routing) - a first-of-its-kind centralized adaptive routing scheme that is applicable in medium size NoCs (dozens of modules). ATDOR dynamically switches between XY and YX routing for each source-destination pair according to globally aggregated traffic loads. Further, we analyze the scalability of large NoCs (hundreds to thousands of modules). We show that despite traffic locality, long-distance (a.k.a. global) packets dominate the bandwidth consumption and the average network delay. To accommodate this phenomenon, we introduce PyraMesh - a novel family of multilevel hierarchical 2D mesh topologies resembling a pyramid structure. In PyraMesh, global packets are separated from the local ones and routed through the upper levels of the hierarchy. The upper hierarchy levels are sparse and therefore provide shorter paths (in terms of hop-distance). Hierarchical architectures utilize, at the upper levels, long links of the order of the die size. RC delays of long links might reach dozens of clock cycles in advanced technology nodes, if delay reduction techniques (e.g. wire sizing and repeater insertion) are not applied. We perform a detailed analysis of the system costs of adjusting long-link delays in hierarchical NoCs to practical target clock frequencies and show that hierarchical NoCs with single cycle links are feasible at present and advanced technology nodes. Finally, we devise a scalable centralized dynamic traffic management scheme for hierarchical NoCs termed DTrD (Dynamic Traffic Distribution). DTrD monitors average buffers occupancy at the upper hierarchy levels using a lightweight off-band infrastructure and dynamically distributes traffic among the hierarchy levels to achieve the lowest average packet delay under both light and heavy traffic loads.