Re: Quad core CPUs loaded at only 50% when running a CPU and mmapintensive multi-threaded task

[Theodore Tso]

On Mon, Aug 25, 2008 at 01:41:17PM +0200, Peter Zijlstra wrote:
> 
> I would certainly consider this for small (< 1M?) files. With mmap the
> faults and pte overhead aren't free either, and the extra memcpy from
> pread() isn't that much.
> 

Even for very big files, if you're only doing a single sequential pass
over a very large file (for example when converting a Canon raw image
file to TIFF format --- I know because I was trying to optimize dcraw
a while aback), you take the page fault for each 4k page, and so
simply using read/pread is faster.  And that's on a single-threded
program.  With a multithreaded program, the locking issues come on top
of that.

Maybe if I had used hugepages it would have been a win, I suppose, but
I never tried the experiment.  And this was several years ago, on much
older hardware, so maybe the relative times of doing the memory copy
versus the page fault, but I wouldn't be surprised if it's even more
expensive to do the mmap, relatively speaking.

							- Ted
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/