On Thu, 24 Mar 2005 14:49:55 -0800 (PST) Christoph Lameter <[EMAIL PROTECTED]> wrote:
> On Thu, 24 Mar 2005, David S. Miller wrote: > > > > prep_zero_page would use a temporal clear for an order 0 page but a > > > nontemporal clear for higher order pages. > > > > That sounds about right to me. > > > > Hmmm, I'm inspired to experiment with this on sparc64 a bit. > > Could you help me fix up this patch replacing the old clear_pages patch? Sure, I'll play with it. Meanwhile, here are some numbers. I changed just the clear_page() implementation on sparc64 so that it used prefetching and normal temporal stores. The machine is a uniprocessor 1.5Ghz Ultra-IIIi, 64K write-through D-cache, 64K I-cache, 1MB L2 cache. I did 4 timed 'vmlinux' builds after a fresh boot: BEFORE: real 9m8.720s user 8m28.345s sys 0m32.734s real 9m2.034s user 8m28.763s sys 0m32.512s real 9m1.848s user 8m28.970s sys 0m32.204s real 9m1.701s user 8m28.715s sys 0m32.394s AFTER: real 9m2.241s user 8m16.633s sys 0m36.451s real 8m53.739s user 8m17.165s sys 0m36.052s real 8m54.089s user 8m17.266s sys 0m36.219s real 8m54.071s user 8m17.473s sys 0m36.073s So, at the very least, my results agree with D. Mosberger's on IA64. At the cost of ~4 seconds of system time, we gain ~11 seconds of user time. I'm pretty much convinced this is a win. I wonder if it matters to do something similar for copy_page*() as well. - To unsubscribe from this list: send the line "unsubscribe linux-kernel" in the body of a message to [EMAIL PROTECTED] More majordomo info at http://vger.kernel.org/majordomo-info.html Please read the FAQ at http://www.tux.org/lkml/
