We’re launching Lyria 3.5 in Google Flow Music, with advances across musicality, lyrics, vocals, and creative control
deepmind.google·6d ago
AI Engineering
Lobsters·29d ago·5 min read
PREEMPT_NONE Is Dead; Your Postgres Probably Doesn’t Care
TL;DR
PostgreSQL performance on Linux 7.0 showed a significant drop in throughput compared to Linux 6.x, raising concerns among users. The issue stemmed from a specific benchmark configuration that was not representative of typical usage scenarios.
✦ Why It Matters
Engineers should ensure proper benchmark configurations to avoid misleading performance assessments.
Key Takeaways
- Review benchmark configurations before drawing conclusions.
- Understand the impact of kernel preemption modes on performance.
- Monitor TLB behavior and page faults in high-load scenarios.
- Stay informed about PostgreSQL performance implications with kernel upgrades.
Full Summary
A recent benchmark from AWS indicated that PostgreSQL's throughput on Linux 7.0 fell to 0.51 times that of Linux 6.x, prompting questions about kernel upgrades. However, this regression is primarily due to a misconfigured benchmark setup, particularly for a system with 96 virtual CPUs and over 100 GB of shared memory.
The underlying issues involve the Linux scheduler, Translation Lookaside Buffer (TLB) behavior, page faults, and a specific spinlock in PostgreSQL that is often overlooked. Lætitia Avrot provided a detailed analysis of this regression, emphasizing that the problem is not inherent to PostgreSQL but rather a result of specific configurations.
The findings suggest that most users can safely upgrade to Linux 7.0 without concern for performance degradation. This situation highlights the importance of proper benchmarking practices when evaluating system performance.