Assessing the computational benefits of AREA-oriented DAG-scheduling

Many modern computing platforms, including "aggressive" multicore architectures, proposed exascale architectures, and many modalities of Internet-based computing are "task hungry"-their performance is enhanced by always having as many tasks eligible for allocation to processors as possible. The AREA-Oriented scheduling (AO-scheduling) paradigm for computations with intertask dependencies-modeled as dags-was developed to address the "hunger" of such platforms, by executing an input dag so as to render tasks eligible for execution quickly. AO-scheduling is a weaker, but more robust, successor to IC-scheduling. The latter renders tasks eligible for execution maximally fast-a goal that is not achievable for many dag s. AO-scheduling coincides with IC-scheduling on dags that admit optimal IC-schedules-and optimal AO-scheduling is possible for all dag s. The computational complexity of optimal AO-scheduling is not yet known; therefore, this goal is replaced here by a multi-phase heuristic that produces optimal AO-schedules for series-parallel dags but possibly suboptimal schedules for general dags. This paper employs simulation experiments to assess the computational benefits of AO-scheduling in a variety of scenarios and on a range of dags whose structure is reminiscent of ones encountered in scientific computing. The experiments pit AO-scheduling against a range of heuristics, from lightweight ones such as FIFO scheduling to computationally more intensive ones that mimic IC-scheduling's local decisions. The observed results indicate that AO-scheduling does enhance the efficiency of task-hungry platforms, by amounts that vary according to the availability patterns of processors and the structure of the dag being executed. © 2011 Springer-Verlag.