THP page cache

Hi,

I believe we will get to transparent huge pages at summit anyway. Below is data points on file-backed transparent huge pages.

The code I have by the time is in my repo[1], see branches thp/pc/v7/p1-base and thp/pc/v7/p2-mmap. Ning Qu works on rebasing shmem/tmpfs support on top of this.

Workloads known to benefit from THP for page cache:

- MongoDB: mongoperf on ramfs shows increase number of iops by 1.9x for r/o and 1.7x for r/w;
- Google search/indexing benchmark shows +3% (in addition to +5% from AnonTHP), on pair with hugetlbfs;
- IOZone shows improvement up to 2.5x on ramfs;

Should help also with:

- Reducing ITLB pressure:
+ x86-64 binaries is ready to be mapped with 2M pages: binutils creates binaries with required file offset and virtual address alignment, no changes required;
+ reported 11% performance increase of RDBMS by putting code to hugetlbfs;
+ MySQL spends 2.5% of cycles in page table walk due ITLB misses[2];
- HPC workloads on many-cores systems (like Xeon Phi): large code and data, small TLB, limited memory bandwidth.

[1] git://git.kernel.org/pub/scm/linux/kernel/git/kas/linux.git
[2] http://research.cs.wisc.edu/multifacet/papers/isca13_direct_segment.pdf

 Kirill A. Shutemov

--
To unsubscribe, send a message with 'unsubscribe linux-mm' in
the body to [email protected].  For more info on Linux MM,
see: http://www.linux-mm.org/ .
Don't email: <a href=mailto:"[email protected]"> [email protected] </a>