I/O layout-aware task scheduler

Abstract

The Non-Uniform Memory Access (NUMA) architecture, containing multiple CPUs, causes non-uniformity in I/O operation. Within a NUMA node, communication between the CPU and I/O devices takes place over Direct Cache Access (DCA), whereas communication between the CPU and I/O devices across distant NUMA nodes takes place over Direct Memory Access (DMA). DMA results in an increased incidence of cache misses during identical network operations, thereby necessitating a greater consumption of CPU cycles. Therefore, utilizing the NIC within other NUMA nodes leads to inefficient usage of CPU resources. To use CPU cycles efficiently for I/O operation, systems require to take into account the non-uniformity in I/O. There are two existing approaches to designing systems that reflect the non-uniformity in I/O: i) The scheduler allocates tasks on the NUMA node where the requested NIC exists. ii) The system extends a NIC connection to the other NUMA nodes so all nodes utilize DCA. However, the first approach does not work for the dynamic workload since it statically allocates resources. The hardware methodology only accommodates a single NIC system, resulting in considerable synchronization overhead arising from simultaneous access by all nodes to a singular NIC.In this work, we design a I/O layout-aware scheduler that handles dynamic workload while preserving the scalability inherent in general-purpose schedulers. Tasks are categorized into local or remote tasks based on where they are requested. Each node preferentially processes local tasks and processes remote tasks if there are no local tasks. To overcome the limitation of static allocation scheduling, we adopt work stealing to achieve work conservation under dynamic workload. Lastly, different NUMA affinitive tasks are mixed in the node when the processor takes remote tasks. To solve this problem, we mark the worker that processes remote tasks. Our scheduling algorithm demonstrates a throughput improvement of up to 48\% in the case of a balanced workload compared to the existing general-purpose scheduler. Our scheduler shows up to 78\% higher throughput for an unbalanced workload than strict boundary scheduling.