William Lee Irwin III wrote:

quoted

c014f9e9 502737   1.64565     .text.lock.file_table
c01b1890 719105   2.35391     __copy_to_user
c0135768 911894   2.98498     __get_page_state
c013ee50 952823   3.11895     do_anonymous_page
c01436d0 1079864  3.53481     page_add_rmap
c01438cc 1186938  3.8853      page_remove_rmap
c0106f38 17763755 58.1476     poll_idle

On Thu, Dec 19, 2002 at 02:31:32AM -0800, Andrew Morton wrote:

Is that improved?

Only in the qualitative sense. No statistically significant differences
in running times are observable.


William Lee Irwin III wrote:

quoted

pfn_to_nid() got lots of icache misses. Try using a macro.

On Thu, Dec 19, 2002 at 02:31:32AM -0800, Andrew Morton wrote:

What's the callsite?

I was not able to collect this information and would be much obliged
to hear of how to do so for instruction cache profiling.


On Thu, Dec 19, 2002 at 02:31:32AM -0800, Andrew Morton wrote:

Actually, just looking at mmzone.h, I have to say "ick".  The
non-NUMA case seems unnecessarily overdone.  eg:
#define page_to_pfn(page)
	((page - page_zone(page)->zone_mem_map) + page_zone(page)->zone_start_pfn)
Ouch.  Why can't we have the good old `page - mem_map' here?

There is no reason why it could not be re-established for the Pee Cee
case. IMHO It would be excellent to have the proper architecture
discrimination defines in place so that Pee Cees could utilize the vastly
simpler calculation. So long as this is devolved to arch code there is no
danger of page->virtual being required by arch-independent code.


Thanks,
Bill
--
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/