Intel’s Linux kernel test robot has detected a significant performance improvement – 3888.9% to be exact – in the mainline Linux kernel, thanks to a single line of code change. The improvement was achieved with the “will-it-scale.per_process_ops” scalability test case on an Intel Xeon Platinum (Cooper Lake) server.
The patch, introduced last week, limits THP alignment of anonymous mappings to PMD-aligned sizes and aims to fix performance regressions in certain workloads. The change only affects one line of code but has had a substantial impact on performance.
Intel’s automated test robot has been monitoring kernel changes for years and has reported positive results in the past. This latest improvement is a significant boost, especially in specialized cases where performance matters.
However, it’s worth noting that this change has also introduced regressions in some workloads, such as cactusBSSN, which experienced up to 600% slowdowns due to TLB or cache aliasing issues caused by the aligned boundaries. The patch fixes these regressions but allows for a more even distribution of mappings.
To address this, the patch adds a condition that mapping sizes must be multiples of PMD size, rather than at least PMD size. This change aims to balance performance gains with minimal disruption to existing workloads.
Source: https://www.phoronix.com/news/Intel-Linux-3888.9-Performance