Paper Title: Improved Flash Performance Using the New Linux Kernel I/O Interface

Paper Abstract: The Linux Kernel has recently added a new I/O interface called io_uring. This brings faster and more efficient I/O operations on Linux when compared to other Linux AIO, which is currently used by various software projects such as Ceph, ScallaDB, SeaStar, and Nginx. All of these utilize Linux AIO to improve I/O performance, but AIO is seen as difficult to use and has limited use cases. Io_uring also solves some problems with Linux AIO. It is also important for software developers and architects to understand the performance characteristics and behavior of io_uring on low latency devices - especially flash and Intel Optane based I/O devices. This talk will showcase performance numbers when using io_uring on NAND and Intel Optane NVMe SSDs. We will go through some performance analysis of io_uring on commonly used storage stacks like disaggregated storage using NVMe-oF and virtualized storage using QEMU/KVM vhost. We will also compare io_uring against other I/O stacks such as Linux AIO and SPDK.

Paper Author: Vishal Verma, Performance Engineer, Intel

Author Bio: Vishal Verma is a performance engineer within Intel's Data Center Group. He has extensive experience in performance analysis and workload characterization of several storage solutions and open-source projects such as Ceph, Swift, and Intel's SPDK. Vishal received his B.Tech in Electronics & Communications from GGSIP University in Delhi, India, and a Masters in EE from Arizona State University.