Escalonamento tolerante a sabotagem em grades computacionais entre-pares.

Abstract

The computational grids are infrastructures to agregate computing power. They evolved in the sense of forming free-to-join communities over the Internet and won the title of peer-to-peer (P2P) grids. It is important to notice that when any user can freely join and leave a system, it becomes more succeptible to damages caused by malicious users. A solution to this problem is to apply sabotage tolerance techniques, generally based on replication, to estimate the computation correctness. Sarmenta has verified that the employment of reputation-based fault tolerance techniques in the task scheduling may promote high confidence levels for the computation results, at the same time that minimizes replication costs, when compared to the tradional voting technique to choose a correct result. The objective of this work is to evaluate the usage of scheduling heuristics that adapt themselves to the machines’ confidence level in P2P grids, applying the fault tolerance technique proposed by Sarmenta. This technique assumes the knowledge of some properties, which are difficult to foresee. The obtained results demonstrate that there are more than one implementation of sabotage-tolerant scheduling. Three scheduling heuristics were evaluated and they present advantages and disadvantages, concluding that, depending on the grid environment, one heuristic might have a better performance than the others.