A COMPREHENSIVE TEST BENCH FOR THE EVALUATION OF SCHEDULING HEURISTICS

Computational problems can usually be divided into a number of (sub)tasks. The data dependencies between these tasks can be described by a so called task graph. Since the task graph scheduling problem is known to be NP-hard, researchers devised an innumerable number of heuristic algorithms to solve this problem best. Up to now, these heuristics were analyzed by using a set of task graph problems, which were rarely revealed to the public. A comparison of different scheduling algorithms presupposes th availability of their implementations. Our contribution c onsists of a test bench with 36000 task graph problems and their optimal solutions with respect to homogeneous target architectures. These test cases are structured according t o several task graph properties and target architectures’ si zes and enable researchers to analyze the overall quality of their heuristics’ results, to detect strengths and weaknes ses of their algorithms and to compare them with those of other researchers. We used the current intermediary version of this test bench to compare some well known scheduling algorithms, namely DLS [1], ETF [2], HLFET [3] and MCP [4]. Surprisingly, the heuristics show a very similar behav ior regarding the quality of the obtained results, exhibiti ng the same strengths and weaknesses, differing only by few percent.