Task and Interrupt Packing for Efficient CPU Utilization

With technological advancements, scale-out approaches utilizing many Central Processing Unit (CPU) cores on general-purpose servers have become common to enhance performance. To reduce initial costs and lower operational costs, efficient CPU utilization with as few cores as possible is required. However, standard schedulers, such as the one in the Linux kernel, do not achieve this. This is because one of the goals of standard schedulers is to allocate execution time fairly among tasks and they tend to assign cores to tasks abundantly. Therefore, in this paper, we verify the hypothesis that CPU utilization efficiency can be optimized by packing tasks and interrupts onto as few cores as possible while considering latency requirements. Specifically, to improve CPU utilization efficiency, we propose and implement a method that combines two approaches on different timescales: task and interrupt packing to minimize the number of CPU cores required, and load-balancing to avoid excessive wait times based on CPU contention states to reduce task latency. The proposed method differs from existing techniques in two key aspects: First, it incorporates interrupts, traditionally treated separately from tasks, into packing calculations based on estimated interrupt processing time. Second, it proactively avoids delay impacts by using the kernel’s task queue length, rather than directly observing task processing time. Performance evaluation demonstrated the feasibility of the proposed method. It packed tasks and simulated interrupts onto a single core while distributing load to another core when task queue length is long. Furthermore, compared to simply packing tasks onto the minimum number of cores, the proposed method successfully reduced maximum latency by 36% (12% under conditions with simulated interrupts). These results showed that it is feasible to efficiently utilize the CPU by packing tasks and interrupts onto as few cores as possible while considering latency requirements.