scheduler changes ??

[Randy.Dunlap]

This is something that I noticed (and Jens noticed :) while I was
working on adding some bounce IO stats to BIO.

Is this purely/mostly scheduler related?
Are these improvements well-known, expected?
I have't heard much, if anything, about them.
They are terrific (in a good sense of the word).

All workloads were run on a VA Linux 4400, 4-proc x86, with 4 GB RAM.
All times are in minutes:seconds.

The workloads are "fillmem" and "mmap002", both from Juan
Quintela's memtest suite.

"fillmem" workload (3 runs with each kernel):
(6 instances of 'fillmem 700')

> 2.4.16 vanilla:          3:02, 2:48, 3:08
> 2.4.16 + block-highmem:  2:51, 2:54, 2:55

> 2.5.2-pre11:             0:30, 0:30, 0:53

And when I use "mmap002" (2 runs with each kernel/options):
(5 instances of 'mmap002 500')

2.4.16 + block-highmem + default idle loop:   3:49, 1:56
2.4.16 + block-highmem + idle=poll:           1:59, 1:59

2.5.2 + default idle loop:                    1:32, 1:37
2.5.2 + idle=poll:                            1:31, 1:37

Kernel profiles on these show that most time is being spent in
default_idle() or poll_idle().

Comments?

Thanks,
  ~Randy