Worst-case Performance Analysis Of3D-NoC Based On Network Calculus

Performance analysis for Network-on-Chip (NoC) is one of the hottest topicsin NoC research area. Most of the previous work focused on either average per-formance or the best-case, but few on the worst-case. With worst-case performance,it can not only reinforce NoC design confidence, but also guarantee NoC quality.Network calculus is one of these tools that can provide effective worst-case per-formance analysis. In NoC, multi-path routing is often preferable than single-pathrouting since it can better balance workload and thus provide better performance.However, performance analysis with multi-path routing is much more difficult dueto complicated contention scenarios. Based on network calculus, this thesis studiesworst-case performance of multi-path minimal routing on3-D mesh NoCs, with thestarting point of2D-NoC. The main contributions are as follows:(1) Basic analysis algorithms on worst-case performance of2D multi-pathminimal routing NoCA per-flow delay bound analysis technique for multi-path routing is firstlypresented. Then a contention matrix is defined to capture NoC contention status.Based on this contention matrix, an effective non-uniform traffic splitting strategyis proposed to improve the worst-case network performance with two differentstrategies: node based and flow based non-uniform traffic splitting. Experimentswith synthetic traffic flows and an industrial case verify the effectiveness of thenon-uniform traffic splitting techniques. Experiments results show that it has amaximum of44.9%improvement of the node based non-uniform traffic splittingtechniques against the full uniform splitting, and a maximum of69.3%improve-ment for the flow based non-uniform traffic splitting. The industrial example showsthat the flow based method achieves a maximum improvement of27.3%against theuniform traffic splitting.(2) Worst-case performance analysis models of3D multi-path minimalrouting NoCBased on network calculus, a traffic splitting technique is presented to trans-form multi-path problem into single-path for3D multi-path minimal routing NoC.Then a sub-flow classification algorithm is proposed to identify and processing thetraffic contention patterns for sub-flows with traffic splitting. Experiments show that in3D multi-path routing NoC, the worst-case performance become worse thanthat of the single-path routing NoC, even when the average case is getting better.(3) Worst-case performance optimization of TSV based3D-NoCA degree based contention matrix is proposed to improve the scalability of theworst-case performance analysis algorithms based on the previous work, which caneffectively reduce the storage space. Two kinds of traffic splitting schemes are alsopresented to optimize the worst-case performance of Throuth Silicon Vias (TSV)based3D-NoC, with special focus on the TSV load balancing. One is full path andthe other is paritail path. Experiments show that the partial path based method ismore effective than the other one in3D NoC worst-case performance optimizing,with a maximum improvement of52.3%against the non-optimized case.