The results in this chapter show the positive impact of M apIO, combined with RCN oC, on
the WCTT values of the core-to-I/O flow leading to solve the problem of dropping Ethernet frames. We have seen that M apIO leads to a significant reduction of 94% for the WCTT values,
in two cases: 1) when including unused cores in the region of the critical applications, and 2) when allocating a less-constrained critical application. However, this mapping has a reasonably limited impact on the core-to-core congestions. Actually, we have shown that there is a positive impact of the presence of unused cores on the core-to-core flows congestions. However, in the case of U Ni = 0, this impact could be negative.
7.5
Conclusion
This chapter presents an evaluation of both of RCN oC and M apIO. We have first shown on the
case study A how RCN oC impacts the WCTT of the core-to-I/O flows, avoiding in some cases
to drop Ethernet frames. However, the reduction of the WCTT values is not sufficient to avoid this problem. Thus, applying M apIO on this case study is a solution to solve this problem and
this by reducing the contention on the core-to-I/O flows.
However, the mapping strategy applied without reducing the pessimism in the computation of the WCTT could not be the best solution. We have illustrated on the case study C the need
to combine RCN oC with M apIO to avoid dropping Ethernet frames. The results show on this
case study that even applying M apIO, the WCTT values computed by the recursive method
leads to drop the Ethernet frame. Despite the small reduction of this WCTT when applying
RCN oC (2.5%), this is sufficient to solve the problem.
Finally, we evaluate the impact of the rules of M apIO and the presence of the unused cores
generated by this strategy on both core-to-I/O and core-to-core flows. Our results show on realistic avionics case studies that the core-to-I/O transmission delays are significantly reduced, up to 94%. Meanwhile, the internal congestion for the core-to-core flows can increase up to 28.8%, but slightly impacting this congestion in the other cases we have considered. Besides, we have noticed on some case studies the inability of existing mapping strategy to allocate all applications whose size does not exceed the size of NoC, unlike M apIO.
Chapter 8
Conclusion
Contents
8.1 Summary of Thesis Contributions . . . 147 8.2 Future Work . . . 149
8.1
Summary of Thesis Contributions
In this thesis, we are interested in the use of NoCs in real-time systems interconnected to sensors and actuators via Ethernet. In this context, a congestion in the NoC, due to wormhole switching, delays the core-to-I/O flow (coming from Ethernet) leading to an overflow of the buffer of the Ethernet interface which is of limited capacity. Therefore, incoming Ethernet frames could be dropped. Real-time packet schedulability analysis must then be done, taking into account all the types of flows. The objective of this thesis was to analyze the WCTT of the different types of flows and to reduce the WCTT of the core-to-I/O flow in order to avoid the drop of Ethernet frames. We have illustrated two main problems in the existing methods when applying them in this context over a Tilera-like architecture:
1. Existing WCTT computing methods do not model the pipeline transmission of flits over wormhole NoCs, thus leading to an over-approximation of the WCTT values. Actually,
these methods consider that an analyzed flow can be blocked by others flows while these last flows have not reach their destinations. Besides, they add the transmission delays of all direct and indirect blocking flows to the transmission delay of the analyzed flow. The pessimism values of the WCTT, especially for the core-to-I/O flows, leads to drop Ethernet frames.
2. Existing contention-aware mapping strategies aim to minimize only the inter-core conges- tion without taking into account the requirements of I/O communications of applications. Then, the WCTT of core-to-I/O flows depends on the congestions generated by applica- tions allocated next to the Ethernet interfaces. These mapping strategies cannot reduce the congestion on the core-to-I/O flows, leading to drop Ethernet frames.
In order to reach our objective, i.e. to avoid the drop of Ethernet frames, two approaches have been proposed:
1. A WCTT computing method, noted RCN oC, that models the pipeline transmission of
flits. For this purpose, we have defined three properties to reduce both the number of scenarios to be explored when performing a WCTT analysis and the computed WCTTs values. Using these properties, we compute the maximal blocking delay a flow can suffer from the blocking flow. This leads to eliminate the need to wait till the blocking flow reaches its destination. Besides, we identify the indirect flows that do not impact the transmission of an analyzed flow. Then, the delay of these flows are not added to the transmission delay of the analyzed flow. We have implemented these properties in an algorithm based on a recursive method to compute the WCTT of the flows.
2. A static mapping strategy of critical and non critical real-time flows, noted M apIO that
reduces the WCTT of core-to-I/O communications over Tilera-like NoC. This mapping is divided into two phases. In the first phase, the NoC is split into regions where critical applications are allocated in priority in dedicated regions close to memory and Ethernet controllers. Thus, the lengths of core-to-I/O communications of critical and non-critical applications are thus reduced. This phase ensures a mapping of all applications without