On Wed, 2004-02-04 at 13:59, Timothy Miller wrote:

Dave McCracken wrote:

quoted

Um, wrong answer.  When you ask for more than one page from the buddy
allocator  (order greater than 0) it always returns physically contiguous
pages.

Also, one of the near-term goals in VM is to be able to allocate and free
large pages from the main memory pools, which requires that something like
order 9 or 10 allocations (based on the architecture) succeed.

What's the x86 large page size?  4M?  16M?  For the sake of arguement, 
let's call it 4M.  Doesn't matter.

2M with PAE on.  4MB with it off.

Let's say this defragmenter allowed the kernel to detect when 1024 4k 
pages were contiguous and aligned properly and could silently replace 
the processor mapping tables so that all of these "pages" would be 
mapped by one TLB entry.  (At such time that some pages need to get 
freed, the VM would silently switch back to the 4k model.)

This would reduce TLB entries for a lot of programs above a certain 
size, and therefore improve peformance.

This is something that would be interesting to pursue, but we already
have hugetlbfs which does this when you need it explicitly.  When you
know that TLB coverage is a problem, that's your fix for now. 

The question is:  How much overhead really is caused by TLB misses?  The 
TLB in the Athlon is like 512 entries.  That means it can know about 2 
megabytes worth of 4k pages at any one time.

I'm not sure how the Athlon works, but the PIII has very few TLB entries
for large pages.  The P4 can put large or small entries anywhere.  

--dave

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