On the scope of applicability of the ETF algorithm

Abstract

Superficially, the Earliest Task First (ETF) heuristic [1] is attractive because it models heterogeneous messages passing through a heterogeneous network. On closer inspection, however, this is precisely the set of circumstances that can cause ETF to produce seriously sub-optimal schedules. In this paper we analyze the scope of applicability of ETF. We show that ETF has a good performance if messages are short and the links are fast and a poor performance otherwise. For the first application we choose the Diamond DAG with unit execution time for each task and the multiprocessor system in the form of the fully connected network. We show that ETF partitions the DAG into lines each of which is scheduled on the same processor. The analysis reveals that if the communication times between pairs of adjacent tasks in a precedence relation are all less than or equal to unit then the schedule is optimal. If the communication time is equal to the processing time needed to evaluate a row then the completion time is 0(^/n) times more than the optimal one for an n x n Diamond DAG. For the second application, we choose the join DAG evaluated by two connected processors.