Re: 2.6.5-rc3-mm4

[Nick Piggin <nickpiggin@yahoo.com.au>]

Andrew Morton wrote:
> markw@osdl.org wrote:
> 
>>I reran DBT-2 to with ext2 and ext3 (in case you were still interested)
>> on my 4-way Xeon system with 60+ drives:
>> 	http://developer.osdl.org/markw/fs/dbt2_project_results.html
>>
>> Aside from the from the drop you're already aware of since 2.6.3, it
>> looks like DBT-2 takes another smaller hit after 2.6.5-rc3-mm2.  Here's
>> a brief summary from the link above:
> 
> 
> The profile is interesting:
> 
> 3671973 poll_idle                                63309.8793
>  77750 __copy_from_user_ll                      637.2951
>  64788 generic_unplug_device                    487.1278
>  62968 DAC960_LP_InterruptHandler               336.7273
>  53908 finish_task_switch                       361.7987
>  52947 __copy_to_user_ll                        441.2250
>  29419 dm_table_unplug_all                      439.0896
>  25947 __make_request                            17.9938
>  18564 dm_table_any_congested                   199.6129
>  13785 update_queue                             104.4318
>  13498 try_to_wake_up                            20.3590
>  12736 __wake_up                                114.7387
>  12560 kmem_cache_alloc                         163.1169
>  12221 .text.lock.sched                          40.7367
> 
> - There's a ton of idle time there.
> 
> - The CPU scheduler is hurting.  Nick and Ingo are patching up a storm to
>   fix a similar problem which Jeremy Higdon is observing at 200,000
>   IOs/sec.  This will get better.

Versus this for 2.6.3:

12825181 poll_idle                                221123.8103
219606 schedule                                 126.7201
194233 __copy_from_user_ll                      1541.5317
191707 __copy_to_user_ll                        1597.5583
149704 DAC960_LP_InterruptHandler               800.5561
120972 generic_unplug_device                    822.9388
  87891 __make_request                            60.9931
  49207 try_to_wake_up                            74.5561
  42647 do_anonymous_page                         65.5100

Which looks like it is taking a lot longer (is it the same test?)
It is difficult to tell how idle each one is due to lack of total
ticks reported, but, copy_to/from_user is 3% the amount of idle
time in 2.6.3, while being 3.5% the amount of idle time in your
profile.

So 2.6.3 could be relatively more idle than -mm.

Do we know what the 2.6.3 regression is caused by? Or is it
likely to be the CPU scheduler?