On 12/02/15 05:08 AM, Tommy Reilly wrote: > We (nuodb) developed a patch to have jemalloc call MADV_NOHUGEPAGE on > all memory it mapped but this only works if THP setting on OS is > [madvise] so its not a complete fix. It does give our customers the > flexibility to use madvise instead of never to work around this problem. > If such a patch would be welcome I will put one together but we > consider it a half fix. Personally I think linux should be changed so > that THP can be disabled by a process regardless of system setting but > that doesn't appear to be possible at the moment.
If it's on 'madvise', then it will only use huge pages if ranges are marked with MADV_HUGEPAGE. The MADV_NOHUGEPAGE hint is a no-op unless the range was already marked with MADV_HUGEPAGE or if it's set to the 'always' setting. It can certainly be disabled by a process regardless of the system setting, but it's only opt-out with 'always' anyway.
signature.asc
Description: OpenPGP digital signature
_______________________________________________ jemalloc-discuss mailing list [email protected] http://www.canonware.com/mailman/listinfo/jemalloc-discuss
