Re: [Starpu-devel] Heft/dmda: locking in compute_all_performance_predictions.

[Cyril Roelandt <cyril.roelandt@inria.fr>]

On 11/22/2012 06:08 PM, Samuel Thibault wrote:
> Cyril Roelandt, le Thu 22 Nov 2012 17:16:36 +0100, a écrit :
> > if (exp_end[worker][nimpl]>  max_exp_end)
> >          max_exp_end = exp_end[worker][nimpl];
> > _STARPU_PTHREAD_MUTEX_UNLOCK(&sched_mutex[worker])
> >
> >
> > The exact same code can be found in compute_all_performance_predictions() in
> > dmda, except that we do not lock the worker's mutex. Why is that ?
>
> I believe it's simply a bug. One should indeed take the lock, to make
> sure to be getting a coherent set of {exp_start, exp_len} values. It
> might be useful to only read those in the critical section, or even turn
> the mutex into an rwlock.

OK. I'll keep using a mutex for now. While we are at it, do we really 
need to keep this:


if (exp_end[worker][nimpl]>  max_exp_end)
         max_exp_end = exp_end[worker][nimpl];

in the critical section ? max_exp_end is a local variable, and exp_end 
is a parameter of the function, I'm not sure they can be simultaneously 
modified by multiple threads.


Cyril Roelandt.