Scheduler Woes: Bisecting Early Performance Regressions Found In Linux 6.19
With Stress-NG just being micro-benchmarks. the other regression I bisected so far on Linux 6.19 was the Nginx HTTPS regression. This benchmark is evaluating the Nginx web server performance on the local host for serving static content via HTTPS. The Nginx regression had many of the same problematic commits as the Stress-NG regression and was also tracked back to being introduced by the kernel scheduler changes.
The Nginx regression is a 9% performance drop on Linux 6.19 with the testing on the AMD Ryzen Threadripper. It wasn't as clear-cut as the Stress-NG regression and thus a few skips over the span of the bisecting. Ultimately though it was tracked down to the kernel scheduler code.
👁 Linux 6.19 regression for Nginx web server
The likely scheduler commit introducing the Nginx regression was sched/fair: Revert max_newidle_lb_cost bump. This commit aimed toa ddress performance regressions in database workloads from a prior commit. The commit did acknowledge regressing the synthetic Schbench scheduler benchmark on the system of Intel engineer Peter Zijlstra but was merged anyhow, It would appear that this commit is also what is harming the Nginx performance on the Linux 6.19 testing I've done thus far.
👁 Linux 6.19 Nginx regression bisected
That's what I have gone through thus far. Depending upon interest level and support I may dig through some other performance issues I've seen on Linux 6.19 Git thus far but for now back on to running other benchmarks that are more likely to make ends meet. Those that appreciate my consistent Linux hardware benchmarking can show their support by joining Phoronix Premium or tips via PayPal and Stripe are also welcome and much appreciated.
If you enjoyed this article consider joining Phoronix Premium to view this site ad-free, multi-page articles on a single page, and other benefits. PayPal or Stripe tips are also graciously accepted. Thanks for your support.