New Linux kernels bring performance improvements
Summary
In 2012 kernels 3.0.39 and 3.2.25 were the first to deliberately bring performance optimisations into the stable and longterm branches. Mel Gorman and Greg Kroah-Hartman had extended the rules to also allow changes that fix important performance or interactivity problems. Typical systems were hardly expected to get faster, but on an SGI UV system with 640 cores, throughput when writing to tmpfs rose from 300 to 430 MByte/s.
Ideas
- Until then, stable kernels had only accepted bug fixes.
- Performance patches carry a higher risk of new bugs.
- Every change has to have arrived in the main branch first.
- Gorman posted the changes on the mailing list for review in advance.
Insights
- Rules for stable branches constantly weigh safety against benefit.
- Large systems benefit from optimisations that cannot be measured on normal computers.
Facts
- The rule change took place during the development of Linux 3.5.
- On the SGI UV system, 120 threads wrote to tmpfs in parallel.
References
Critique
- The report only gives measurements for a special system; figures for ordinary servers are missing.
Recommendations
- With stable updates, watch whether performance changes are included as well, and test them.
- Measure performance improvements on your own hardware and workload.
Links to the original source and the Web Archive open in a new tab.