Research Of Temperature Prediction And Task Allocation In MPSoC Based On 3D NoC

With the development of integrated circuit technology, multiprocessor system-on-chip (MPSoC) have already been employed as the main design of the next generation of single-chip processors. As a result of the insufficiency of the traditional global interconnection, which restrict the improvement of system and the expansion of the dimension on NoC architectures. The design of MPSoC using Network-on-Chip (NoC) structure will be the main orientation of study on MPSoC in the future, it provides the larger interconnection bandwidth, lower network power consumption and higher transistor density than the traditional design to achieve higher performance. However, the higher transistor density and the increase operating frequency of the processors have caused overheated temperature in the chip. Then, the inevitable heat dissipation problem become one of the key issues to improve the performance of the MPSoC. This thesis focuses on the heat dissipation issues of NoC-based MPSoC, meanwhile, a thorough study has been made on the temperature prediction and task allocation to solve the issues on the optimization of prediction accuracy and the extension of the prediction time length, the optimization of the information sharing strategy, and the optimization of the peak temperature and the heat distribution.Based on the thermal resistance and capacitance (Thermal RC) model, the thesis firstly proposes a predictive thermal model combined with the second derivative, starting from the nonlinear feature of temperature variation. The model predicts the temperature accurately with low complexity, and increases the prediction time length within a certain prediction error range to reduce the number of times that the prediction module is invoked and the extra power consumption. Experimental results show that, compared to the existing model combined with the first derivative, the proposed model increases the prediction length by 1.6 times within the same acceptable prediction error range. When the prediction time length is increased to 2.5s, the prediction accuracy of the proposed model is 3.84% higher than that of the contrastive model.Aiming at the strategy of information sharing, an equilibrium partitioning multicast transmission method is proposed, which can obtain a trade-off between the startup latency and the network latency. The optimal partitioning number can be determined by this method for the current source node, and then, the corresponding areas are divided according to the location of the source node to make the total latency as small as possible. Experimental results show that, compared to the existing recursive partitioning (RP) transmission method, the proposed strategy reduces the total latency by 16.4% within the same network configuration at most. When the packet length is 1 flit,5 flits and 15 flits, the total latency of the proposed method is 2.4%,4.1% and 18.3% lower than RP transmission method respectively within three kinds of mesh topology (4x4x3,8x8x3 and 16x16x3) at most.In order to balance the temperature generated by the processors, and reduce the peak temperature in the system, the thesis proposes a task set allocation (TSA) scheme firstly that can balance the regional temperature. Then, an improved task set allocation scheme based on the top power (TP-TSA) is presented aiming to reduce the peak temperature. The improved TP-TSA scheme can reduce the peak temperature in the system, and reduce the likelihood of hot spot. Experimental results show that, compared to the existing Adapt3D scheme, the TP-TSA scheme reduces the likelihood of hot spot by 37.43%, and increases the throughput by 30.44% on average. Besides, the likelihood of reducing the hot spot and the throughput can be increased by 47.88%,46.35% at most within three kinds of mesh topology (2×2×2,2×2×3 and 2x2x4).