Scaling ZFS for NVMe - Allan Jude - EuroBSDcon 2022

> Now that we have NAND storage, would we create filesystems differently? Are there ways to make a step-change in overall performance somewhere? you might enjoy Allan Jude's "Scaling ZFS for NVMe" talk [0] from 2022. the tl;dw is that modern NVMe drives have gotten so fast that filesystem developers are having to revisit assumptions that have been true for decades about which operations will be slow vs. fast. flash storage that exposed a SATA interface and pretended to be a superbly fast hard drive was an evolutionary improvement of sorts. NVMe is more of a revolutionary improvement, it does away with the spinning-rust assumptions. for example, SATA limits a device to one command queue with 32 commands; NVMe makes it almost unlimited (64k queues with up to 64k commands) [1] there is also a corner of the NVMe spec [2] that allows you to treat an SSD directly as key-value storage instead of block storage with a filesystem implementing a key-value layer on top of it. this is promising but AFAIK no mainstream filesystem is using it yet, and I'm not sure if support has even trickled down into consumer drives or if it's market-segmented off for datacenter use only. 0: [this video] 1: [link] 2: [link]

At EuroBSDcon 2022, Allan Jude gave the presentation "Scaling ZFS for NVMe": > Learn about how ZFS is being adapted to the ways the rules of storage are being changed by NVMe. In the past, storage was slow relative to the CPU so requests were preprocessed, sorted, and coalesced to improve performance. Modern NVMe is so low latency that we must avoid as much of this preprocessing as possible to maintain the performance this new storage paradigm has to offer. > An overview of the work Klara has done to improve performance of multiple ZFS pools of large numbers of NVMe disks on high thread count machines. […] > A walkthrough of how we improved performance from 3 GB/sec to over 7 GB/sec of writes. * [this video] When ZFS was created spinning rust was still the main thing, and SSDs were gaining popularity, so ZFS created "hybrid storage" pools: * [link] * [link] * [link] * [link] * [link] Still useful for use cases that lean towards bulk storage (versus IOps).