A heuristic based real time task assignment algorithm for heterogeneous multiprocessors

Abstract

In this paper a Heuristic Based Improved Linearly Decreasing Weight Particle Swarm Optimization algorithm (ILDW-PSO) is proposed to solve real time task assignment in heterogeneous processors. To achieve better quality solution to the particles, existing LDW-PSO encompasses to ILDW-PSO, which includes Genetic algorithm mutation operator. The objective is to minimize the maximum utilization of the processors with energy aware load balance condition which reduces the energy consumption. Experimental results show that ILDW-PSO outperforms the existing PSO algorithms interms of maximum utilization and normalized energy consumption for both consistent utilization matrix and inconsistent utilization matrix.

CPU-GPU heterogeneous implementations of depth-based foreground detection

“…For further explanation, the per-video-frame execution times (unit: msec) of deviceCPU and deviceGPU on platform1 are shown in Fig. 4, and speedups of the data decomposition-based CPU-only (i.e., all tasks were executed on deviceCPU), GPU-only (i.e., all tasks were executed on deviceGPU), CPU-GPU (i.e., workload was distributed into both deviceCPU and deviceGPU based on the workload ratio computed for each task [7][8][9]) method, and the proposed method are compared in Fig. 5.…”

“…5. Speedup comparison between the data decomposition-based method such as [7][8][9] and the proposed method.…”

“…6(c). As explained, the more numbers of tasks, the more numbers of workload ratios should be determined in the data decomposition-based heterogeneous method [7][8][9]. Furthermore, for example, Task5 in the next version of the target vision system (i.e., CLAHE [12] for contrast enhancement) matches well with deviceCPU, rather than deviceGPU, because Task5 is also an irregularly structured computation such as Task4.…”

“…However, the goal of this study is not to optimize a specific vision task. Instead, we optimize the entire vision application (i.e., depth-based foreground detection) with data decomposition-based parallel solutions for each vision task such as [7][8][9]. By measuring the speed of each task in a computing device (i.e., deviceCPU or deviceGPU) through the OpenCL-based implementation, however, we can determine an appropriate computing device for each parallelized task whose computational characteristics are different.…”

“…We can then reduce the total execution time further by balancing the execution times of the deviceCPU and deviceGPU with the proposed pipeline scheduling algorithm. Note that many data decomposition-based parallel methods for a regularly-structured task have been reported on a CPU-GPU heterogeneous computing platform [7][8][9]. To the best of our knowledge, however, this is the first study that detects pigs at night on CPU-GPU heterogeneous computing platforms with a pipelined execution of the foreground detection consisting of several regularly-and irregularly-structured tasks.…”