On Thu, May 26, 2011 at 01:13:58PM -0400, Eric B Munson wrote:
> +#define YES "yes"
> +#define NO "no"
Upstream 2.6.39 was changed to 0/1, that differs from 2.6.38 :(. It's
probably better to support only 0/1 and let it fail on 2.6.38.
> +#define TRANS_ENABLE "/sys/kernel/mm/transparent_hugepage/enabled"
This sounds ok if you want hugeadm to be able to tweak everything
(maybe set madvise mode and then use THP only in a few apps with
madvise through --trans-madvise).
> +#define TRANS_DEFRAG "/sys/kernel/mm/transparent_hugepage/defrag"
> +#define KHUGE_DEFRAG
> "/sys/kernel/mm/transparent_hugepage/khugepaged/defrag"
I'd drop these two, they're too much for debugging, in libhugetlbfs
context especially I can't see anybody using THP with defrag off that
turns off compaction and reclaim entirely and just polls the buddy
allocator. This is more for debugging only, not useful to tweak.
> +#define KHUGE_SCAN_PAGES
> "/sys/kernel/mm/transparent_hugepage/khugepaged/pages_to_scan"
> +#define KHUGE_SCAN_SLEEP
> "/sys/kernel/mm/transparent_hugepage/khugepaged/scan_sleep_millisecs"
> +#define KHUGE_ALLOC_SLEEP
> "/sys/kernel/mm/transparent_hugepage/khugepaged/alloc_sleep_millisecs"
This is useful too. These three plus the "/enabled" are the only 4 I'd
consider worth tweaking not only for debugging.
> @@ -97,6 +109,20 @@ void print_usage()
> CONT("Adjust pool 'size' lower bound");
> OPTION("--obey-mempolicy", "Obey the NUMA memory policy when");
> CONT("adjusting the pool 'size' lower bound");
> + OPTION("--trans-always", "Enable transparent huge pages always");
> + OPTION("--trans-madvise", "Enable transparent huge pages with madvise");
> + OPTION("--trans-never", "Disable transparent huge pages");
> + OPTION("--trans-defrag-always", "Don't limit THP defrag");
> + OPTION("--trans-defrag-madvise", "Limit THP defrag to madvise areas
> only");
> + OPTION("--trans-defrag-never", "Disable defrag during THP allocation");
> + OPTION("--khugepaged-defrag-off", "Disable defrag during THP
> promotion");
> + OPTION("--khugepaged-defrag-on", "Enable defrag during THP promotion");
> + OPTION("--khugepaged-pages <pages to scan>", "Number of pages that
> khugepaged");
> + CONT("should scan on each pass");
> + OPTION("--khugepaged-scan-sleep <milliseconds>", "Time in ms to sleep
> between");
> + CONT("khugepaged passes");
> + OPTION("--khugepages-alloc-sleep <milliseconds>", "Time in ms for
> khugepaged");
> + CONT("to wait if there was a huge page allocation failure");
Wouldn't it be cleaner to use a "thp" prefix instead of "trans". And
add the thp prefix to thp-khugepaged too? (but the latter is ok also
keeping it khugepaged I guess, I'd still s/trans/thp/)
> +.B --khugepaged-pages <pages to scan>
> +
> +Configure the number of pages that khugepaged should scan on each pass
> +
> +.TP
> +.B --khugepaged-scan-sleep <milliseconds>
> +
> +Configure how many milliseconds khugepaged should wait between passes
> +
> +.TP
> +.B --khugepages-alloc-sleep <milliseconds>
> +
> +Configure how many milliseconds khugepaged should wait after failing to
> +allocate a huge page to throttle the next attempt.
> +
> +.PP
> The following options affect the verbosity of libhugetlbfs.
Good description.
Thanks,
Andrea
------------------------------------------------------------------------------
vRanger cuts backup time in half-while increasing security.
With the market-leading solution for virtual backup and recovery,
you get blazing-fast, flexible, and affordable data protection.
Download your free trial now.
http://p.sf.net/sfu/quest-d2dcopy1
_______________________________________________
Libhugetlbfs-devel mailing list
Libhugetlbfs-devel@lists.sourceforge.net
https://lists.sourceforge.net/lists/listinfo/libhugetlbfs-devel
