July 8, 2013
Participants: Kirill A. Shutemov, Jan Kara, and Aneesh Kumar K.V.
People tagged: (none)
Kirill A. Shutemov suggested a discussion of extending transparent hugepages from anonymous pages to also support huge pages in the page cache. Jan Kara asked if dirtiness was to be tracked on a partial-hugepage basis, as is currently done in some cases for buffer heads, or if the idea was instead to split hugepages in the partial-dirty case. Kirill replied that he intended to avoid the overhead of splitting by using huge pages only if the file was open read-only, thus deferring any dirty-page issues. Jan asked what performance benefits would accrue in real workloads, and Kirill agreed to bring numbers into the next LKML discussion.
Aneesh Kumar K.V also indicated an interest and asked after architectures (like 64-bit powerpc) where block size was less than the page size, along with page-fault implications for the mmap() case.