Re: [BENCHMARK] 12% improvement in fork load from mm3 to mm4

[Andrew Morton]

[e-mail] wrote:
>
> > autoconf build test creates about 1.2 million processes in
> > 35 minutes.  On quad xeon, there was an improvement of
> > ~ 12% between 2.5.70-mm3 and 2.5.70-mm4.
> > 
> >               Avg of 3 runs
> > 2.5.70-mm3    829.0
> > 2.5.70-mm4    737.9
> 
> profile shows the biggest change is in do_page_fault.

Well that's amusing.  In mm4 I replaced the patch which runs
remap_file_pages() against all prot_exec mappings with a patch which just
prefaults those mappings into pagecache.

The difference _should_ be that in mm3, all pte's are set up at mmap time.
So -mm4 should have more pagefaults, not less.

Something fishy is going on there.  Makes one wonder if remap_file_pages()
is successfully avoiding the minor fault.