Re: [Starpu-devel] StarPu: performance model doesn't work

[Cedric Augonnet <cedric.augonnet@inria.fr>]

Hello,

Just to clarify why StarPU has such a behaviour: imagine that we assume that 
a model is calibrated as soon as you have one measurement, if the kernel is 
not perfectly regular or when there is an extra overhead for the first 
invokation of the kernel (that sometimes happen with library that are 
initialized lazily for instance), then you might have a prediction that is 
not accurate and you potentially end up by never assigning a task to some 
type of processing unit.

In src/starpu_parameters.h you would find this:
#define STARPU_CALIBRATION_MINIMUM 10

So the current threshhold is 10 measurements to consider that the model is 
calibrated. This value is of course arbitrary, but as you can imagine, the 
more samples, the better the accuracy!

I don't know if this is very clear ...

Best,
Cédric

Le 13 juin 2011 à 16:59, Leo Li a écrit :

> Hello, Cedric
> 
> Thanks for your response.
> 
> Yes, I have obtained one measurement for each task. So what is the
> minim number of samples needed when considering the model?
> 
> Yes, I run with the STARPU_CALIBRATE flag set to 1 when I observe this
> weird behaviour.
> 
> Please see attached my test case, I developed it based on vector scale
> example. In the README file I wrote about how I run the test case.
> 
> Thank you for your help.
> 
> 
> On Mon, Jun 13, 2011 at 11:24 AM, Cedric Augonnet
> <cedric.augonnet@inria.fr> wrote:
>> Hello Lu Li,
>> 
>> It looks like you only have obtained one measurement for each task? The 
>> heft/dmda policies try to have some minimum number of samples before 
>> considering that the model is calibrated...
> 
>> 
>> Are you still running with the STARPU_CALIBRATE flag set to 1 when you 
>> observe this weird behaviour?
>> 
>> Perhaps you can give us your test-case and specify the different 
>> environment  in case it is simple enough to be shared?
>> 
>> Thanks for your report!
>> 
>> Cedric
>> Le 13 juin 2011 à 10:44, Leo Li a écrit :
>> 
>>> Hi, Everybody
>>> 
>>> My name is Lu Li. I am a master student at Linkoping University, Sweden.
>>> 
>>> I currently do my master thesis based on StarPU under professor
>>> Christoph W. Kessler. And I got a problem, could someone help me.
>>> Thanks.
>>> 
>>> My problem is the performance model doesn't help the scheduling. After
>>> calibration, I get data shown in attachment1. From the data we can see
>>> the scheduler should pick CPU platform to run. Actually what I
>>> observed was that the schedulers always picked CUDA. I have tried
>>> heft, dmda, dm schedulers, and they all pick CUDA. It seems they don't
>>> take performance data into account. Anybody who can help me?
>>> 
>>> Thanks in advance.
>>> 
>>> --
>>> Faithful
>>> Lu Li
>>> 
>>> 
>>> Contact: +46704759692
>>> <performance_data.jpg>_______________________________________________
>>> Starpu-devel mailing list
>>> Starpu-devel@lists.gforge.inria.fr
>>> http://lists.gforge.inria.fr/cgi-bin/mailman/listinfo/starpu-devel
>> 
>> 
> 
> 
> 
> -- 
> Faithful
> Lu Li
> 
> 
> Contact: +46704759692
> <performanceModel.tar>