Synthetic Traffic Analysis

5.2 Synthetic Traffic Analysis

In synthetic traffic analysis, the physical layer of our simulation environment comprises of 8×8 nodes connected as a mesh array. Each packet has a constant length of 16 flits. The tests performed sent packets at varying injection rates and varying GS packets percentages under different traffics for 25000 cycles. Four types of synthetic traffic patterns were run: uniform, regional, transpose, and HotSpot. In uniform traffic, a node will be the destination of every other nodes with equal probability based on the injection rate. In regional, 90% of the packets are sent to the destination at distance less than 3 hops, and most transmission occur among two neighbor routers. In transpose traffic, a node at coordinate (i,j) will sent a packet to the destination which coordinate is (j,i). In HotSpot traffic, most of packets are send to the same destination at once. The traffic is similar with real-case traffic because in SoC most IP have communication with the main CPU. The GS percentage defines the GS packets count versus total packet count, this value not only represents the GS packets utilization but also shows that the architectures can get fit in different QoS requirements.

We analyze the simulation results obtained from low GS percentage first under the uniform traffic. Figs. 5-1(a) and (c) represent the GS packet latency with GS

is lower than BiNoC, and Pre-Req BiNoC has also latency improvement compared with the traditional BiNoC.

0

0

Fig. 5-1. GS Packet Latency at Low GS Percentage under Uniform Traffic.

Figure 5-2 shows the latency of uniform traffic under higher GS percentage from 0.1 to 0.5. We only focus on the latencies from zero-load to saturation which is double of the zero-load latency. The latency of AQ-BiNoC is 15% less than BiNoC even in GS with a percentage of 50%. The simple Pre-Req BiNoC still has similar latency curve comparing with BiNoC. The minor improvement comes from the inter-router arbitration mechanism.

0

0

Fig. 5-2. GS Packet Latency at High GS Percentage under Uniform Traffic.

The following graphs show the BE packet latencies in different QoS percentages over different injection rates. We can notice that the latencies of BE packet in BiNoC, Pre-Req BiNoC and AQ-BiNoC almost have no difference. The reason is that most functions implemented are used to serve GS packets only. When the GS packets travel much fast in the network, the resources occupied by GS packets could also be released as soon as possible. That is why in some cases the BE latency in AQ-BiNoC is better than the one in BiNoC.

0

0

0

Fig. 5-3. BE Latency vs. Injection Rate in GS Percentages from 0.01 to 0.5.

From Figs. 5-4 and 5-5, we can observe that the improvement of packet is not oblivious under regional traffic, the most important reason for the latency of each architecture being such close is that the regional traffic is transmitting data to close destination mostly. The latency decrease is from the anticipative bi-direction channel control only in case where there is no long enough distance of packet transmitting. It causes the Pre-Req BiNoC and AQ-BiNoC have the same performance curves. And they both improve the neighborhood transmission by the inter-router channel control.

0

0

Fig. 5-4. Latency of GS Packet in Regional Traffic.

0

0

Fig. 5-5. Latency of BE Packet in Regional Traffic.

The results of latency of GS packets in transpose traffic are illustrated in Fig. 5-6.

The latency improvement is decreased as the GS percentage grows. The latency of AQ-BiNoC is worse than the original BiNoC as the latency grows up to twice of the zero-load latency. The reason of this effect may be that since the transpose traffic causes congestion at router coordinate (i,i), the penetration ability is limited by half of

shown in Fig. 5-7. The Pre-Req BiNoC has similar behavior with BiNoC. Even in high GS percentage, it also can achieve lower latency than AQ BiNoC without losing one virtual channel buffer transform into penetration buffer.

0

0

Fig. 5-6. Latency of GS Packet in Transpose Traffic.

0

0

Fig. 5-7. Latency of BE Packet in Transpose Traffic.

The saturation behavior in HotSpot traffic is worse than other traffics because under high utilization at small area, a resource of single router will be consumed out at low injection rate. We can discover under the burst traffic AQ-BiNoC still got better latency even in high GS percentage.

0

0

Fig. 5-8. Latency of GS Packet in HotSpot Traffic.

0

0

Fig. 5-9. Latency of BE Packet in HotSpot Traffic.

respectively. Even under the condition where the GS percentage and the injection rate are high, we can observe that the AQ-BiNoC has better latency performance than BiNoC and NoC architectures under uniform traffic.

Fig. 5-10. Latency of AQ-BiNoC vs. BiNoC in Uniform Traffic.

Fig. 5-11. Latency of AQ-BiNoC vs. NoC in Uniform Traffic.

Analyzing the overall performance improvement of uniform traffic as shown in Fig. 5-12, we can find that under higher GS packet percentage, the improvement scale becomes smaller. It is reasonable because performance will become worse as same resources like penetration buffers are shared by more requesters. But the improvement is still effective under higher GS percentage in uniform traffic. This phenomenon can be realized by the characteristic of uniform traffic where the packet flows are separated into whole network equally. Thus, the resources can be shared averagely to whole GS packets which need high transmission quality.

Fig. 5-12. Improved Percentage of AQ-BiNoC vs. BiNoC in Uniform Traffic.

Figures 5-13 and 5-14 illustrate the latencies in regional traffic of AQ-BiNoC vs.

BiNoC and NoC, respectively. Even under the condition where the GS percentage and the injection rate are high, we can observe that the AQ-BiNoC has better latency performance than BiNoC and NoC architectures. The regional traffic sends packets to the destinations which are close to the source coordinates. The improvement is mainly provided by the inter router channel arbitration functionality.

Fig. 5-13. Latency of AQ-BiNoC vs. BiNoC in Regional Traffic.

Fig. 5-14. Latency of AQ-BiNoC vs. NoC in Regional Traffic.

traffic. But we can observe that under higher GS packet percentage, the improvement still does not decay critically.

Fig. 5-15. Improved Percentage of AQ-BiNoC vs. BiNoC in Regional Traffic.

Figures 5-16 and 5-17 show the latencies in transpose traffic of AQ-BiNoC vs.

BiNoC and NoC, respectively. The performance becomes worse when the GS percentage and the injection rate are high, but AQ-BiNoC has better latency performance than BiNoC and NoC architectures under low injection rate. Since transpose traffic sends packets to the destinations which coordinate is the transpose value of the sources, most of the packet flow will stuck and against each other at the routers which coordinate is (i,i) . Comparing with BiNoC, the AQ-BiNoC router gets

less sources; thus, the congestion will be much serious than the competition in BiNoC.

Fig. 5-16. Latency of AQ-BiNoC vs. BiNoC in Transpose Traffic.

Fig. 5-17. Latency of AQ-BiNoC vs. NoC in Transpose Traffic.

Figure 5-18 emphasizes this phenomenon by the worse latency under high GS percentage and high injection rate.

Fig. 5-18. Improved Percentage of AQ-BiNoC vs. BiNoC in Transpose Traffic.

Figures 5-19 and 5-20 show the latencies in hotspot traffic of AQ-BiNoC vs.

BiNoC and NoC, respectively. The overall performance shows that our proposed AQ-BiNoC works well under such similar to real traffic case.

Fig. 5-19. Latency of AQ-BiNoC vs. BiNoC in HotSpot Traffic.

Fig. 5-20. Latency of AQ-BiNoC vs. NoC in HotSpot Traffic.

Figure 5-21 shows the improved percentage of AQ-BiNoC vs. BiNoC. The performance decayed when injection rate goes high. But in overall situation, AQ-BiNoC gets better latency.

Fig. 5-21. Improved Percentage of AQ-BiNoC vs. NoC in HotSpot Traffic.