TechPrufer
Get started
Cloudflare
Cloudflare logoCloudflare · Feature trail

io_uring worker pool management

1 post · Feb 2022

What this trail covers

Cloudflare's io_uring implementation has evolved to provide detailed insights into worker pool management. This includes understanding the distinction between bounded and unbounded workers, how their sizes are controlled by system limits (e.g., RLIMIT_NPROC) and ring buffer configurations, and how to monitor their creation and behavior using kernel tracepoints. The platform now offers granular control and visibility into the underlying runtime for asynchronous I/O processing, enabling developers to optimize performance and resource utilization for network-intensive workloads.

All Cloudflare trails
  1. Missing Manuals - io_uring worker poolThis post introduces the internal workings of io_uring's worker pool management, specifically differentiating between bounded and unbounded workers based on I/O type (regular files/block devices vs. network sockets/special files). It details how unbounded worker pool size is limited by RLIMIT_NPROC and how bounded workers are a function of SQ ring size and CPU count. The post demonstrates how to observe worker creation using kernel tracepoints like `io_uring:io_uring_submit_sqe` and `io_uring:io_uring_poll_arm`, and explains that worker threads are only spawned when I/O requests are explicitly marked as asynchronous (`IOSQE_ASYNC`) or when blocking I/O is necessary, rather than relying on non-blocking polling for sockets by default. It provides a test workload and analysis using `perf` and `bpftrace` to illustrate these concepts.