A cache-aware DAG scheduling method on multicores: Exploiting node affinity and deferred executions

Abstract

With increasingly complex functionalities being implemented in emerging applications, multicores are widely adopted with a layered cache hierarchy, and Directed Acyclic Graphs (DAGs) are commonly employed to model the execution dependencies between tasks. For such systems, scheduling methods can be designed to effectively leverage the cache to accelerate the system execution. However, the traditional methods either do not consider DAGs, or rely on sophisticated static analysis to produce fixed scheduling solutions that require additional hardware support (e.g., cache partitioning and colouring), which undermines both the applicability and flexibility of these methods. Recently, an online cache-aware DAG scheduling method has been presented that schedules DAGs using an execution time model with caching effects considered, eliminating the need for static analysis and additional hardware support. However, this method relies on simple heuristics with limited considerations on both the allocatable cores and the competition between nodes, resulting in intensive inter-node contention that undermines cache performance. This paper proposes CADE, a cache-aware scheduling method for DAG tasks that leverages the cache to reduce DAG makespan. To achieve this, an affinity-aware priority assignment is first constructed that mitigates the competition among nodes for their preferred cores to hit the cache. Then, a contention-aware allocation mechanism is constructed, which (i) accounts for the impact of an allocation decision on the speed-up of other nodes; and (ii) includes the busy cores for allocation by enabling the deferred execution, effectively enhancing the cache performance to accelerate the DAG execution. Experiments show that compared to the state-of-the-art, the CADE significantly reduces the DAG makespan by 24.02% on average (up to 33%) with the cache miss rate reduced by 22.06% on average.