TechPrufer
Get started
PlanetScale
PlanetScale logoPlanetScale · Feature trail

Postgres I/O Method Benchmarking

1 post · Oct 2025

What this trail covers

This post benchmarks the new I/O method configurations in PostgreSQL 18 (`worker` and `io_uring`) against the default `sync` method from PostgreSQL 17. It analyzes performance across various EC2 instance types with different storage configurations (gp3, io2, local NVMe) and concurrency levels using the `sysbench oltp_read_only` workload. The findings highlight that local NVMe drives offer superior price-performance, and while `io_uring` shows promise, `worker` often provides a better balance of performance and flexibility, especially on network-attached storage.

All PlanetScale trails
  1. Benchmarking Postgres 17 vs 18This post introduces and benchmarks the new `io_method` configuration options (`worker`, `io_uring`) in PostgreSQL 18 against the traditional `sync` method. It details a comprehensive benchmarking methodology using `sysbench oltp_read_only` across four EC2 instance types with varying storage (gp3, io2, NVMe) and concurrency levels (1, 10, 50 connections). The analysis quantifies performance differences in QPS and identifies scenarios where `io_uring` excels (high concurrency, CPU-bound workloads) and where `worker` or even `sync` might be preferable, particularly on network-attached storage. It also discusses the cost-performance implications of different storage solutions and provides insights into why `io_uring` may not always be the winner, referencing external analysis on index scans and checksum bottlenecks.