Transparent Hugepage Performance On Linux 6.18 LTS: Madvise vs. Always

Written by Michael Larabel in Memory on 6 January 2026 at 10:45 AM EST. Page 2 of 4. 24 Comments.

👁 NAMD benchmark with settings of Input: STMV with 1,066,628 Atoms. always was the fastest.

With workloads like the NAMD molecular biology software, the always THP mode provided around a 6% performance advantage on this AMD EPYC Turin server.

👁 QuantLib benchmark with settings of Size: S. madvise was the fastest.

Meanwhile running in madvise mode allowed for around 3% better performance than the always THP mode for the QuantLib quantitative finance software package.

👁 NWChem benchmark with settings of Input: C240 Buckyball. always was the fastest.

👁 Xcompact3d Incompact3d benchmark with settings of Input: Turbulent Boundary Layer. madvise was the fastest.

And in some workloads such as where the process is automatically opting into their desired THP configuration, there ends up being no measurable difference.

👁 OpenFOAM benchmark with settings of Input: drivaerFastback, Small Mesh Size, Mesh Time. always was the fastest.

👁 OpenFOAM benchmark with settings of Input: drivaerFastback, Small Mesh Size, Execution Time. always was the fastest.

👁 OpenFOAM benchmark with settings of Input: drivaerFastback, Medium Mesh Size, Mesh Time. always was the fastest.

👁 OpenFOAM benchmark with settings of Input: drivaerFastback, Medium Mesh Size, Execution Time. always was the fastest.

👁 OpenFOAM benchmark with settings of Input: drivaerFastback, Large Mesh Size, Mesh Time. always was the fastest.

👁 OpenFOAM benchmark with settings of Input: drivaerFastback, Large Mesh Size, Execution Time. always was the fastest.

The always THP mode was beneficial to the OpenFOAM CFD software.

24 Comments - Next Page

Page: 1 2 3 4 Next Page

URL: https://www.phoronix.com/review/thp-madvise-always/2

⇱ Transparent Hugepage Performance On Linux 6.18 LTS: Madvise vs. Always Review - Phoronix

Transparent Hugepage Performance On Linux 6.18 LTS: Madvise vs. Always