More Benchmarks Of The Performance Pullback In Linux 4.20

Last night I published some benchmarks after finding Linux 4.20 is regressing in several workloads compared to Linux 4.18/4.19 and at least was affecting Intel Core i9 "HEDT" boxes. Here are more affected workloads regressing on Linux 4.20 and it's not just limited to high-end hardware.

This morning I decided to check in on my automated bi-daily kernel benchmarks on LinuxBenchmarking.com. It's all automated and thus don't necessarily have the time to look at the data too often (even though PTS' LinuxBenchmarking.com does also provide email notifications when auto-detecting possible regressions), but in looking back at the archived data it too captured a significant performance pullback on multiple systems on Linux 4.20.

This data does show the performance regression occurred right at the start of the Linux 4.20 cycle. It also showed regressions occurring with an Intel Core i3 7100, Xeon E5-1680 v3, and Core i7 7740X. But a Core i5 8400 and Core i5 7600K along with AMD CPUs didn't appear impacted.

Even the very basic Smallpt C path tracer showed a performance regression, most notably on the Core i3 7100 and Core i7 7740X hardware.

Most regressions seemed to have been with multi-threaded workloads, but even the PHP performance saw a major setback on many of the Intel systems under test. That pullback came at the very start of the Linux 4.20 cycle and has been reproduced ever since. (For the systems only with data on select days, it's just what happens when the systems hung or have a part failure until I have the time to investigate and repair... The bi-daily automated Linux kernel benchmarks are done without any project funding or support from any organization, nor ads on that page, thus is basically just a hobby project and for PTS/OpenBenchmarking/Phoromatic demonstration purposes.)

The timed Apache compilation showed a small regression in its build times. The longer-running compiler tests from yesterday show this issue much more pronounced.

HMMer appears to be another affected workload.

Fortunately, with the performance regression being very evident on the powerful Core i9 7980XE and the Phoronix Test Suite doing all the heavy-lifting, the issue is currently being bisected... Should have the data available in the next few hours for what is causing this measurable drop in performance.

Update: The bisected answer.