[Starpu-devel] Suggestions about creating starpu threads

[Fangli Pi <hpcfapix@hlrs.de>]

Hi Andra,

I have created a new email as I think the only open point now is the point 10.

Regarding to your suggestion about creating StarPU threads and get power data 
directly from mf_agent, I have discussed with my colleague and got an 
impression that it may take significant work, while the architecture of our 
monitoring framework would need to be changed and I am not sure if the new 
method will be efficient as before (for example, rapl need to be started and 
stopped every time you need the power measurements).

Besides, although there are overhead of mf_server query, it will not effect 
the performance of StarPU tasks. Since the query only happens when the user 
turns on the training mode of power model (like calibrate for the performance 
model), which is not the real runtime mode of StarPU. After enough training, 
the power model could be used to predict the power consumption of the task 
during runtime and for scheduling. In another word, it may become slow once 
the user turns on the training mode, however, for all the other cases 
performance will not be effected. After all, the performance during training 
is not so important, right?

Please let me know if it would make sense. Or if you think it is very 
important to keep the performance in a training mode, please feel free to 
correct me.

Best regards,
Fangli
-- 
High Performance Computing Center (HLRS)   
University of Stuttgart                    
Nobelstr. 19                               
D-70569 Stuttgart, Germany

phone: ++49-711-685-60442