Re: [GIT] Experimental threaded udev

[Alan Jenkins <alan-jenkins@tuffmail.co.uk>]

Kay Sievers wrote:
> On Thu, May 28, 2009 at 16:35, Alan Jenkins <alan-jenkins@tuffmail.co.uk> wrote:
>   
>> Now available for your delight and/or horror.
>>
>> <http://github.com/sourcejedi/udev/commits/threading-v0.3>
>>
>> For now, I'm still treating this as a patch series.  That is, I may
>> publish future versions with a rewritten history.  I'll preserve the old
>> branches though.
>>
>> It turns out the MADV_DONTFORK hack I was so proud of is
>> implementation-dependant, i.e. a dirty hack.  However, I'm confident
>> that glibc can and should be modified to do it for all programs.  And it
>> is so worth it.  On my test machine, threading alone goes from 2s
>> boot-time coldplug to 1.3-ish.  MADV_DONTFORK takes it down to 0.7-ish.
>> The hack is contained in the last patch, "when forking a program, only
>> copy the stack of the _current_ thread".
>>     
>
> Is that a single or dual CPU box?
>
> With the threaded version, it's 0.18 (1.51 -> 1.33) seconds faster
> here for a full coldplug run on a:
> Intel(R) Core(TM)2 Duo CPU U9400  @ 1.40GHz.
>
> It might be that the threaded version will only behave that much
> better on a single CPU machine?
>   

No, it's not that.  I'm afraid the result of ~0.7 seconds result was an 
accident; the bootchart looks suspicious and it didn't obtain when I 
retested.  The "fixed" version of the last patch doesn't work, and I 
don't think it can.

The version that crashed on boot may be useful for comparison purposes 
though.  I will try and see how much I can reduce the page fault 
overhead without using threads.  Maybe just recycling the event 
processes would bring similar gains, with less of the risks of threads.

Alan