[Gllug] Large Page Patch

[Dave Jones]

On Tue, Aug 13, 2002 at 05:54:01PM +0100, Huw Lynes wrote:
 > Just wondering if anyone here has any experience with the large page 
 > patch. I was talking to an engineer today who swears blind that the 
 > performance problems we are having with a IDE-FC raid array are down to 
 > linux(ia32) using 64k pages. Don't know if it's true but I suppose it's 
 > worth a try.

ia32 uses 4KB pages by default. Other option is 4MB.
(There's also a funky 2MB option iirc, but there's some catch I don't recall)
I'd be surprised if this patch alone caused that much of a performance
improvement on something I/O bound.

Using large pages just buys you better usage of the CPUs TLB's, which
depending on your workload may not be an improvement at all.
Essentially, you want to be operating on 4MB chunks of data, with
re-use of already-read data. If you're doing streaming I/O, you'll
likely only use those pages once (read it, process it, throw it away).

The 64k number is a 'soft pagesize' that I believe MS Windows uses.
Moving to soft pages may be a happy-compromise giving the best of
both worlds, but again, I'd be surprised on an I/O bound load
to see that much difference.  Soft pages are iirc just clusters of
4KB pages. So if we take a pagefault, we fault in 16 pages instead
of 1.

I'm not a mm person, so take this with a small pinch of salt,
but I'm fairly confident my numbers are right.

        Dave

-- 
| Dave Jones.        http://www.codemonkey.org.uk
| SuSE Labs

-- 
Gllug mailing list  -  Gllug at linux.co.uk
http://list.ftech.net/mailman/listinfo/gllug