Results for BAHIA. For the bursty traffic scenario, the gNoCsim network simulator was used for the evaluations. In the case of BAHIA, the configuration consists of a 2D mesh with 64 switches arranged in a 8 × 8 distri- bution where each node is attached to a switch. Regarding to the traffic pattern generated, each node generates a 0.2 flit/cycle baseline/background traffic following an uniform pattern and, when simulation reaches cycle 10000, traffic bursts are sent to 4 different nodes following a 4-to-1 strategy (each hotspot node will receive traffic from 4 nodes) at 1 flit/cycle until cycle 20000 is reached. Meanwhile, the rest of nodes will continue sending at a 0.2 flits/cycle. See Table I for all configu- ration parameters. The table shows the upper and lower thresholds used to detect the bursty traffic. Parameter Value Topology 8x8 2D regular mesh Virtual networks no BAHIA 2VN BAHIA 1 default VN + 1 extra VN Packet switching VCT Flit size 5 bytes Packet size 10 flits Message size 40 flits BAHIA Upper Limit 0.7 flits/cycle Lower Limit 0.2 flits/cycle Polling interval 500 cycles Notification delay 1 cycle Table 3: Simulation configuration for BAHIA and BAHIA-2 mechanisms.
Results for BAHIA. × For the bursty traffic scenario, the gNoCsim network simulator was used for the evaluations. In the case of BAHIA, the configuration consists of a 2D mesh with 64 switches arranged in a 8 8 distri- bution where each node is attached to a switch. Regarding to the traffic pattern generated, each node generates a 0.2 flit/cycle baseline/background traffic following an uniform pattern and, when simulation reaches cycle 10000, traffic bursts are sent to 4 different nodes following a 4-to-1 strategy (each hotspot node will receive traffic from 4 nodes) at 1 flit/cycle until cycle 20000 is reached. Meanwhile, the rest of nodes will continue sending at a 0.2 flits/cycle. See Table I for all configu- ration parameters. The table shows the upper and lower thresholds used to detect the bursty traffic. Parameter Value Topology 8x8 2D regular mesh Virtual networks no BAHIA 2VN BAHIA 1 default VN + 1 extra VN Packet switching VCT Flit size 5 bytes Packet size 10 flits Message size 40 flits BAHIA Upper Limit 0.7 flits/cycle Lower Limit 0.2 flits/cycle Polling interval 500 cycles Notification delay 1 cycle Table 3: Simulation configuration for BAHIA and BAHIA-2 mechanisms. In Figure 17.(a) we see the accepted network traffic. As can be seen, the simulation begins with a background traffic of 0.2 flits/cycle approximately. However, starting at transient 20 (each transient is made of 500 cycles, thus being cycle 10000), four nodes receive bursty traffic, each from four different nodes. Nodes receiving the burst are not able to dispatch bursty traffic thus contention appears. In no-BAHIA NoC both virtual networks are affected equally since there is no traffic separation, either bursty traffic and normal traffic is allocated without distinction into any virtual network so performance degradation appears in both virtual networks due to HoL blocking effects. Nevertheless, in the case of BAHIA NoC, when bursty traffic is detected at the end nodes, sender nodes are notified so this bursty traffic is isolated into the extra virtual network so there is no HoL blocking in the default virtual network, improving performance in the overall network. Average network latency, shown in Figure 17.(b), behaves similarly, achieving best performance for the implementation with BAHIA. In this case, non-burst traffic latency keeps roughly unaltered. In Figure 18 a study of the robustness of BAHIA is shown. For these simulations the delay of the notification has been varied between 1 (value in ...
