LWN.net Logo

Huge pages part 3: Administration

Huge pages part 3: Administration

Posted Mar 4, 2010 17:32 UTC (Thu) by mel (subscriber, #5484)
In reply to: Huge pages part 3: Administration by nix
Parent article: Huge pages part 3: Administration

> I wonder how much of this will be rendered less important by transparent hugepages, when
> they're accepted? Will hugetlbfs ever be considered a legacy thing for the sake of those rare
> apps that need it?

Not much of it is rendered unimportant in the event transparent support
gets merged. Transparent support solves one particular problem well but it is far from covering
all the bases. There are some important limitations to transparent support including;

1. backing text/data with huge pages (important for scientific apps written in Fortran).
Whatever about backing data, backing text would require significant changes to how the
page cache lookup is implemented in Linux which is done lightly
2. shared memory (databases being the biggest consumer)
3. Using huge pages beyond what the page allocator provide - 1G on AMD or 16G on powerpc
4. Not all architectures can implement transparent huge page support. X86 32-bit because of
a page flag limitation (could be worked around), powerpc has limitations on where huge
pages can be placed (technically could be worked around but it would be of immense
difficulty), IA-64 would need major rearchitecting to change its pagetable format with
corresponding changes to the core before it could even consider transparent support.
Only X86-64 works right now but conceivably SPARC and sh support could be added
without too much hair loss

Call me biased, but the stuff covered in this articles will be important for a long time yet.


(Log in to post comments)

Huge pages part 3: Administration

Posted Mar 5, 2010 0:18 UTC (Fri) by nix (subscriber, #2304) [Link]

I see what you mean. I hadn't spotted that the current transparent
hugepage support was restricted to data pages, and item 4 is a killer.
(Well, it's a killer as long as Linux cares about 'minority'
architectures, which I hope it always will, although the really old ones
like sparc32 and voyager do eventually die off because there's nobody left
who cares about them).

Copyright © 2013, Eklektix, Inc.
Comments and public postings are copyrighted by their creators.
Linux is a registered trademark of Linus Torvalds