Framework for simulation of fault tolerant heterogeneous multiprocessor system-on-chip

Due to the ever growing requirement in high performance data computation, current Uniprocessor systems fall short of hand to meet critical real-time performance demands in (i) high throughput (ii) faster processing time (iii) low power consumption (iv) design cost and time-to-market factors and more importantly (v) fault tolerant processing. Shifting the design trend to MPSOCs is a work-around to meet these challenges. However, developing efficient fault tolerant task scheduling and mapping techniques requires optimized algorithms that consider the various scenarios in Multiprocessor environments. Several works have been done in the past few years which proposed simulation based frameworks for scheduling and mapping strategies that considered homogenous systems and error avoidance techniques. However, most of these works inadequately describe today\u27s MPSOC trend because they were focused on the network domain and didn\u27t consider heterogeneous systems with fault tolerant capabilities; In order to address these issues, this work proposes (i) a performance driven scheduling algorithm (PD SA) based on simulated annealing technique (ii) an optimized Homogenous-Workload-Distribution (HWD) Multiprocessor task mapping algorithm which considers the dynamic workload on processors and (iii) a dynamic Fault Tolerant (FT) scheduling/mapping algorithm to employ robust application processing system. The implementation was accompanied by a heterogeneous Multiprocessor system simulation framework developed in systemC/C++. The proposed framework reads user data, set the architecture, execute input task graph and finally generate performance variables. This framework alleviates previous work issues with respect to (i) architectural flexibility in number-of-processors, processor types and topology (ii) optimized scheduling and mapping strategies and (iii) fault-tolerant processing capability focusing more on the computational domain; A set of random as well as application specific STG benchmark suites were run on the simulator to evaluate and verify the performance of the proposed algorithms. The simulations were carried out for (i) scheduling policy evaluation (ii) fault tolerant evaluation (iii) topology evaluation (iv) Number of processor evaluation (v) Mapping policy evaluation and (vi) Processor Type evaluation. The results showed that PD scheduling algorithm showed marginally better performance than EDF with respect to utilization, Execution-Time and Power factors. The dynamic Fault Tolerant implementation showed to be a viable and efficient strategy to meet real-time constraints without posing significant system performance degradation. Torus topology gave better performance than Tile with respect to task completion time and power factors. Executing highly heterogeneous Tasks showed higher power consumption and execution time. Finally, increasing the number of processors showed a decrease in average Utilization but improved task completion time and power consumption; Based on the simulation results, the system designer can compare tradeoffs between a various design choices with respect to the performance requirement specifications. In general, designing an optimized Multiprocessor scheduling and mapping strategy with added fault tolerant capability will enable to develop efficient Multiprocessor systems which meet future performance goal requirements. This is the substance of this work