QoS-Aware, Cost-Efficient Scheduling for Data-Intensive DAGs in Multi-Tier Computing Environment

In today’s scientific landscape, Directed Acyclic Graphs (DAGs) are pivotal for representing task dependencies in data-intensive applications. Traditionally, two dominant bottom-up DAG scheduling approaches exist: one overlooks communication contention and the other fails to exploit parallelization for improving latency. This study distinguishes itself by advocating a top-down approach prioritizing latency or cost optimization in multi-tier environments to fulfill QoS and SLA requirements. Our strategy effectively mitigates bandwidth contention and facilitates parallel executions, leading to substantial completion time reductions. Our findings suggest that myopic knowledge-based scheduling, emphasizing latency or cost minimization, can yield benefits comparable to its look-ahead counterparts. Through latency-efficient and cost-efficient topological sorting, our <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">wDAGSplit</i> strategy introduces a two-stage partitioning and scheduling approach. Its simplicity and adaptability extend its usability to DAGs of any scale. Evaluated on over 100,000 real-world DAG applications, <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">wDAGSplit</i> demonstrates latency enhancements of up to 80x compared to Edge-only scenarios, 15x to Near-Edge-only, and 6x to Cloud-only. In terms of cost, our approach achieves enhancements of up to 60x compared to Edge-only scenarios, 250x to NE-only, and 70x to Cloud-only. Moreover, for DAGs with 50 tasks, we achieve 5x reduced latency compared to previous approaches, along with a complexity reduction of up to 24 times.