Current Proceedings on Technology

Yazarlar: Babak Nariman Jahan, Habib Izadkhah

Konular:-

Anahtar Kelimeler:Social: Supercompiler,Task Graph,Evolutionary Approach,Multi-Processor

Özet: Nowadays, using several processors to achieve high computational speed is regarded as a principal in most of the scientific applications. Meanwhile, a Supercompiler is used to convert sequential code to parallel code to be run on multi-processor systems. One of the most important issues about supercompiler is parallelization of the nested iteration loops, which are frequently used in scientific computations. Here, there are two complex and effective issues; the first is the distinction of dependencies among the various iterations and conversion to the independent iterations and the second is scheduling of independent iterations to run on multi-processor systems. In this paper, each independent iteration is regarded as a task and the task graph is used to show the dependency among the dependable iterations. As the scheduling of tasks graph is difficult and complex, many attempts have been made to find a solution that is close to the optimum using the evolutionary process. Therefore, we have clustered the task graph using the grading formula and solved the clusters using evolutionary algorithm, including self-repairing, fine and reward (guidance). The characteristics of the proposed algorithms include the increase of speed to arrive at a solution, cost reduction of communications among the processors and load balancing between the processors. Finally, the practical results obtained from the implementation of the proposed method indicate that one can find an appropriate and stable schedule within a far less period than the similar algorithms.

---