Optimizing for NVMe Drives - The 10 Microsecond Challenge

Dive into the world of NVMe drive optimization in this 50-minute conference talk by Stefan Hajnoczi from Red Hat. Explore the "10 Microsecond Challenge" and learn about latency numbers, IOPS, and architectural considerations for maximizing NVMe performance. Discover key concepts such as IO request lifecycle, completion mechanisms, and PCIe device assignment. Gain insights into configuration techniques using libvert, CPU idle halt poll, and iopoll. Compare bare metal vs NVMe performance, and understand the role of Virtio Block and IOThreads. Delve into advanced topics like userspace NVMe drivers, polling queues, and Linux IO Poll. Examine the QMU Storage daemon and explore future directions in NVMe optimization. Walk away with a comprehensive understanding of strategies to tackle the challenges of optimizing for ultra-fast NVMe drives in modern computing environments.

Introduction
NVMe Drives
Latency Numbers
IOPS and Latency
Rethink the architecture
IO request lifecycle
Latency
Completion mechanisms
PCIe device assignment
Configuration
libvert
Example
CPU Idle Halt Poll
Results
iopoll
Bare metal vs NVMe
Virtio Block
IOThreads
IOThreads Configuration
Userspace NVMe Driver
Polling Queues
AIO Fast Path
Linux IO Poll
QMU Storage daemon
Future direction
Summary
Conclusion