An Efficient Scheduling Algorithm for Multi-mode Tasks on Near-Data Processing SSDs

Near-Data Processing (NDP) architectures have been proposed to alleviate the large overhead of data movement between the host and the Computational Storage Device (CSD) by offloading tasks to the CSD. In NDP architectures, each task can run in multiple modes according to the resource it takes for computing, such as the CPU of the host, the accelerator or the processor of the CSD. However, existing task scheduling algorithms on NDP architectures are unaware of the multi-mode tasks, leading to increased completion time of tasks and low resource utilization. In this paper, we propose a Multi-Mode Task Scheduling (MMTS) algorithm to optimize the completion time of the multi-mode tasks in NDP architectures. MMTS employs a greedy strategy to fully use the computing resources in the host and the CSD and align the completion time of the tasks by picking the proper modes. Our experimental results show that MMTS achieves 20.6% performance improvement on average over the state-of-the-art task scheduling algorithm on NDP-based system.