Why TensorFlow Lite Has Been Running Slower On Recent Linux Kernels

The culprit was narrowed down to sched/fair: Remove meaningless imbalance calculation.

Unfortunately, while the commit title may appear mundane, this sched/fair work is part of a much larger set of changes introduced back for Linux 5.5. In fact, the patch is part of the work covered last year on Phoronix and titled Linux 5.5's Scheduler Sees A Load Balancing Rework For Better Perf But Risks Regressions. This scheduler balancing rework is what appears to be to blame for the lower TensorFlow Lite performance we are seeing on the kernels over the past year.

That rework was quite an intrusive change last year and in the kernels since there has obviously only been further scheduler changes. With post-5.5 kernels we have seen some improvements made but still not to the pre-5.5 levels from the testing we have been doing on some of the more powerful boxes. Thus for now there is no easy/straightforward solution given the complexity of the change, thus moving on to looking at other active kernel regressions...

If you appreciate the relentless, daily Linux benchmarking on Phoronix you can show your support by joining Phoronix Premium for ad-free access and multi-page articles on a single page. PayPal tips are also welcome. Back to benchmarking and looking at some of the other open issues.