NVSwap: Latency-Aware Paging using Non-Volatile Main Memory

Page relocation (paging) from DRAM to swap devices is an important task of a virtual memory system in operating systems. Existing Linux paging mechanisms have two main deficiencies: (1) they may incur a high I/O latency due to write interference on solid-state disks and aggressive memory page reclaiming rate under high memory pressure and (2) they do not provide predictable latency bound for latency-sensitive applications because they cannot control the allocation of system resources among concurrent processes sharing swap devices. In this paper, we present the design and implementation of a latency-aware paging mechanism called NVSwap. It supports a hybrid swap space using both regular secondary storage devices (e.g., solid-state disks) and non-volatile main memory (NVMM). The design is more cost-effective than using only NVMM as swap spaces. Furthermore, NVSwap uses NVMM as a persistent paging buffer to serve the page-out requests and hide the latency of paging between the regular swap device and DRAM. It supports in-situ paging for pages in the persistent paging buffer avoiding the slow I/O path. Finally, NVSwap allows users to specify latency bounds for individual processes or a group of related processes and enforces the bounds by dynamically controlling the resource allocation of NVMM and page reclaiming rate in memory among scheduling units. We have implemented a prototype of NVSwap in the Linux kernel-4.4.241 based on Intel Optane DIMMs. Our results demonstrate that NVSwap reduces paging latency by up to 99% and provides performance guarantee and isolation among concurrent applications sharing swap devices.