A private file descriptor per process

2026-03-11 · Source: Hussein Nasser · Field: Technology & Digital — Software Development & Engineering, Cloud Computing & IT Infrastructure · Depth: Advanced, quick

Summary

The discussion centers on optimal placement for IO_uring file descriptors in backend processes. Two primary design choices were considered: assigning a single private IO_uring instance to each backend process or utilizing a shared pool of IO_uring instances. The private instance approach offers simplicity and avoids contention but risks exhausting file descriptors and increasing memory usage, especially in systems like PostgreSQL with many connections (e.g., 100 default). Conversely, a shared pool reduces file descriptor count but introduces significant contention, necessitating mutexes and leading to higher latency and complexity. Ultimately, the shared pool model was abandoned due to its severe contention issues, with the private IO_uring instance per backend process being the preferred and final design choice.

Key takeaway

For system architects designing high-performance I/O systems, particularly those involving many concurrent backend processes like database servers, you should prioritize minimizing contention over reducing file descriptor count. While a shared IO_uring pool might seem efficient, the overhead of managing mutexes and resolving contention can introduce more latency and complexity than simply allocating a private IO_uring instance per process. Evaluate the true cost of contention versus resource duplication in your specific environment.

Key insights

Contention overhead in shared IO_uring pools outweighs the cost of private file descriptors per process.

Principles

• Private resources simplify concurrency.
• Shared resources introduce contention.
• Mutex overhead can exceed resource duplication.

Topics

• IO_uring
• File Descriptors
• Backend Processes
• Concurrency Management
• System Design
• Contention

Best for: Software Engineer, AI Architect, Research Scientist

Related on AIssential

• The Path of a Request: A Tour of Modern Web Architecture — Modern web architecture uses layered systems to funnel traffic, optimizing performance by handling requests at the earliest possible stage.
• Someone is always getting blocked — True asynchronous I/O offloads blocking tasks to a lower layer, like the kernel, which still performs synchronous operations.
• Tech industry averages just 5% GPU utilization, report finds — Widespread GPU and CPU overprovisioning leads to significant financial waste in tech infrastructure.
• The Invisible Layer: Why the success of your data pipeline starts in the File System — Understanding file systems is crucial for optimizing data pipelines and diagnosing performance issues in data engineering.
• Scaling Recommendation Systems with Request-Level Deduplication — Request-level deduplication is a cross-cutting technique that optimizes storage, training, and serving for large-scale recommendation systems.
• Pinterest Engineers Eliminate CPU Zombies to Resolve Production Bottlenecks — Unused default agents can leak kernel state, causing hidden CPU bottlenecks and system instability.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Hussein Nasser.