Results for OCRL mechanism. × To evaluate OCRL we use real application traffic. The system configuration implements a tile- based system like in Figure 1, but with a 10 10 2-D mesh, and each tile is composed of a processor core, a private L1 cache, a bank of a L2 shared cache, a memory directory bank to be used with the directory-based MOESI cache coherency protocol, and 4 memory controllers. The simulation parameters are identical to the ones used for the CMP congestion study presented in Section 2. In order to evaluate the real traffic scenario we launched a set of evaluations with 32 concurrent applications with 3 different placements of the congestion thresholds at the input buffers as shown in Figure 14a, which stands for OCRLA, OCRLB, and OCRLC labels at the figures. In Figure 15 we can see the effect of sharing NoC resources with other applications on the same chip. Please note that the variation between different runs of the same application at the same location is insignificant, so the baseline for this figure is the execution time of just 1 instance closest to one of the memory controllers. As it is seen, there is an average of 1.5% reduction in the execution time for the 3 variants when using the congestion time. Notice that this reduction can make a difference in the total execution time of the system while improving chip utilization if we aggregate all the averaged time reductions. Real-time constraints that influence certain traffic types can benefit also for every reduction. Figure 15: Execution time for 32 concurrent different applications We performed evaluations for all the application types, and similar patterns of congestion problems appear for every application type. We only present the results for the ocean workload. In Figure 16 we show the results for network throughput. The point series labeled as injected reflect the traffic that the network interfaces try to push into the network, and the one labeled as accepted reflects the traffic that is currently accepted and is being forwarded through the network with no congestion control. As shown in the figure, there is an important drop due to congestion between the cycle 130k and cycle 250k. All OCRL variants are able to alleviate the congestion drop by 15% in average, although OCRLC seems to perform better. The variant OCRLD is OCRLC in which we modeled the latency of a faster control network than the one that was set up for previous evaluations. As it can be seen, OCRLD has almost zero congestion and is a demons...
