The efficient scheduling of multi-task jobs across multiprocessor systems has become increasingly critical with the rapid expansion of computational systems. This challenge, known as Multiprocessor Multitask Scheduling (MPMS), is essential for optimizing the performance and scalability of applications in fields such as cloud computing and deep learning. In this paper, we study the MPMS problem under both deterministic and stochastic models, where each job is composed of multiple tasks and can only be completed when all its tasks are finished. We introduce $\mathsf{NP}$-$\mathsf{SRPT}$, a non-preemptive variant of the Shortest Remaining Processing Time (SRPT) algorithm, designed to accommodate scenarios with non-preemptive tasks. Our algorithm achieves a competitive ratio of $\ln \alpha + \beta + 1$ for minimizing response time, where $\alpha$ represents the ratio of the largest to the smallest job workload, and $\beta$ captures the ratio of the largest non-preemptive task workload to the smallest job workload. We further establish that this competitive ratio is order-optimal when the number of processors is fixed. For stochastic systems modeled as M/G/N queues, where job arrivals follow a Poisson process and task workloads are drawn from a general distribution, we prove that $\mathsf{NP}$-$\mathsf{SRPT}$ achieves asymptotically optimal mean response time as the traffic intensity $\rho$ approaches $1$, assuming the task size distribution has finite support. Moreover, the asymptotic optimality extends to cases with infinite task size distributions under mild probabilistic assumptions, including the standard M/M/N model. Experimental results validate the effectiveness of $\mathsf{NP}$-$\mathsf{SRPT}$, demonstrating its asymptotic optimality in both theoretical and practical settings.
翻译:暂无翻译