Re: get_exec_dcookie overhead

[dada1 <da...@co...>]

> Eric,
>
> Do you have an example of the conditions that this change made a
> difference and how much a difference this made on performance?
>
> -Will

Well, I now build the huge program with a special loader script to put the
.text portion at the very begining of vm space...

So I dont remember if I actually tried the patch... it was a long time ago
:)


>
> John Levon wrote:
> > On Thu, Jun 12, 2003 at 02:06:29PM +0200, dada1 wrote:
> >
> >
> >>I wanted to give you this info/hint, but if you feel another kernel guy
> >>could take the job, I can resend the question on the mailing list.
> >
> >
> > I've Cc:ed the mailing list then.
> >
> > Everyone: Eric is seeing significant costs in get_exec_dcookie due to
> > thread stacks being mapped before the EXECUTABLE mmap.
> >
> >
> >>Another point : If a context switch occurs between 2 threads sharing the
> >>same mm_struct, could we optimize sync_buffer() :
> >>
> >>                               /* userspace context switch */
> >>                                new = (struct task_struct *)s->event;
> >>
> >>                                release_mm(mm);
> >>                                mm = take_tasks_mm(new);
> >>
> >>                                cookie = get_exec_dcookie(mm);
> >>with :
> >>
> >>                               /* userspace context switch */
> >>                                new = (struct task_struct *)s->event;
> >>
> >>                                {
> >>                                struct mm_struct *new_mm =
take_tasks_mm(new);
> >>                                if (new_mm != mm)
> >>                                    cookie = get_exec_dcookie(new_mm);
> >>                                release_mm(mm);
> >>                                mm = new_mm ;
> >>                                }
> >>
> >>This change should be easy to do (no modification on the mm_struct), and
> >>certainly could solve my problem (I have only one instance of the
dserver on
> >>my machines)
> >
> >
> > Yes, I think we could definitely do this. Can you please do some
> > comparative tests ?
> >
> > I like the simplicity of it.
> >
> > regards,
> > john
> >
> >
> > -------------------------------------------------------
> > This SF.NET email is sponsored by: eBay
> > Great deals on office technology -- on eBay now! Click here:
> > http://adfarm.mediaplex.com/ad/ck/711-11697-6916-5
> > _______________________________________________
> > oprofile-list mailing list
> > opr...@li...
> > https://lists.sourceforge.net/lists/listinfo/oprofile-list
> >
>