On Sunday 06 June 2004 17:39, you wrote:
> This is an update of the scheduler policy mechanism rewrite using the
> infrastructure of the current O(1) scheduler. Slight changes from the
> original design require a detailed description. The change to the original
> design has enabled all known corner cases to be abolished and cpu
> distribution to be much better maintained. It has proven to be stable in my
> testing and is ready for more widespread public testing now.
>
>
> Aims:
>  - Interactive by design rather than have interactivity bolted on.
>  - Good scalability.
>  - Simple predictable design.
>  - Maintain appropriate cpu distribution and fairness.
>  - Low scheduling latency for normal policy tasks.
>  - Low overhead.
>  - Making renicing processes actually matter for CPU distribution (nice 0
> gets 20 times what nice +20 gets)
>  - Resistant to priority inversion
>  - More forgiving of poor locking
>  - Tunable for a server workload or computational tasks
>
>
> Description:
>  - All tasks start at a dynamic priority based on their nice value. They
> run for one RR_INTERVAL (nominally set to 10ms) and then drop one stair
> (priority). If they sleep before running again they get to run for 2
> intervals before being demoted a priority and so on until they get all
> their intervals at their best priority: 20 intervals for nice 0; 1 interval
> for nice +19.
>
>
> - The staircase elevation mechanism can be disabled and all tasks can
> simply descend stairs using the sysctl:
>
> echo 0 > /proc/sys/kernel/interactive
>
> this has the effect of maintaining cpu distribution much more strictly
> according to nice whereas the default mechanism allows bursts of cpu by
> interactive tasks before settling to appropriate cpu distribution.
>
>
> - The time tasks are cpu bound can be increased by using the sysctl:
>
> echo 1 > /proc/sys/kernel/compute
>
> which extends the RR_INTERVAL to 100ms and disables the staircase elevation
> improving conditions for pure computational tasks by optimising cache
> benefits and decreasing context switching (gains another 1.5% on
> kernbench).
>
>
> Performance:
> - All cpu throughput benchmarks show equivalent or better performance than
> mainline. Note that disabling the interactive setting actually _worsens_
> some benchmarks because of their dependence on yield() so I don't recommend
> disabling it unless you do a comparison first.
> - Interactivity is approximately equivalent to mainline 2.6 but with faster
> application startup and no known behavioural quirks.
>
>
> Comments and testing welcome.
>
>  fs/proc/array.c        |    2
>  include/linux/sched.h  |    8
>  include/linux/sysctl.h |    2
>  kernel/sched.c         |  676
> +++++++++++++------------------------------------
>  kernel/sysctl.c        |   16 +
>  5 files changed, 212 insertions(+), 492 deletions(-)
>
> Can be downloaded here:
> http://ck.kolivas.org/patches/2.6/2.6.7-rc2/patch-2.6.7-rc2-s6.3
>
> and below
> Con
The same patch modified to apply clean on 2.6-7-rc2-mm2.
-- 
        Jan