From: Nick Piggin <piggin@cyberone.com.au>

This one gathers better statistics about the new process problem.  It
improves estimation for initial process IO.  That is, better calculations
for whether it will be worth waiting after a process submits its first
read.

This is done with a per queue average thinktime and seek time for a second
read submitted from a process.

When combined with 3/3, numbers are around the same as mm1 for most long
lived tasks, but much better for things like the top 4 benchmarks.

Probably wants rwhron and the OSDL database guys to give it some testing.

test                                    2.6.0-test9-mm1  2.6.0-test9-mm1-np
Cat kernel source during seq read       0:26.89          0:24.75
Cat kernel source during seq write      9:17.80          0:23.48
ls -lr kernel source during seq read    0:11.03          0:14.68
ls -lr kernel source during seq write   0:49.95          0:08.06

contest no_load                         143s 0 loads     144s 0 loads
contest io_load                         193s 40.2 loads  193s 40.1 loads
contest read_load                       186s 11.6 loads  190s 10.5 loads
contest list_load                       201s 5.0 loads   200s 5.0 loads

pgbench 1 client                        31.3TPS          31.5TPS
pgbench 4 clients                       37.7TPS          37.5TPS
pgbench 16 clients                      42.1TPS          48.1TPS