Re: AIM7 40% regression with 2.6.26-rc1

["Zhang, Yanmin" <yanmin_zhang@linux.intel.com>]

On Tue, 2008-05-06 at 19:39 +0200, Ingo Molnar wrote:
> * Andrew Morton <akpm@linux-foundation.org> wrote:
> 
> > Finally: how come we regressed by swapping the semaphore 
> > implementation anyway?  We went from one sleeping lock implementation 
> > to another - I'd have expected performance to be pretty much the same.
> i.e. we'll always keep yet another task in flight. This can mask wakeup 
> latencies especially when it takes time.
> 
> The patch (hack) below tries to emulate this weirdness - it 'kicks' 
> another task as well and keeps it busy. Most of the time this just 
> causes extra scheduling, but if AIM7 is _just_ saturating the number of 
> CPUs, it might make a difference. Yanmin, does the patch below make any 
> difference to the AIM7 results?
I tested it on my 8-core stoakley and the result is 12% worse than the one of
pure 2.6.26-rc1.

-yanmin

> 
> ( it would be useful data to get a meaningful context switch trace from 
>   the whole regressed workload, and compare it to a context switch trace 
>   with the revert added. )
> 
> 	Ingo
> 
> ---
>  kernel/semaphore.c |   10 ++++++++++
>  1 file changed, 10 insertions(+)
> 
> Index: linux/kernel/semaphore.c
> ===================================================================
> --- linux.orig/kernel/semaphore.c
> +++ linux/kernel/semaphore.c
> @@ -261,4 +261,14 @@ static noinline void __sched __up(struct
>  	list_del(&waiter->list);
>  	waiter->up = 1;
>  	wake_up_process(waiter->task);
> +
> +	if (likely(list_empty(&sem->wait_list)))
> +		return;
> +	/*
> +	 * Opportunistically wake up another task as well but do not
> +	 * remove it from the list:
> +	 */
> +	waiter = list_first_entry(&sem->wait_list,
> +				  struct semaphore_waiter, list);
> +	wake_up_process(waiter->task);
>  }