Abstract:
|
Developing software for exascale systems will become even
more challenging than for today’s systems. Methods for evaluating
the performance of applications and identifying potential
weaknesses are essential for reaching optimal performance.
Though the tools available today are not widely used,
and generally require some expert knowledge.
In recent years different deep learning techniques have enjoyed
great success in various fields, and especially in image
recognition. Though it is still to find its way in to the area of
application performance evaluation.
This work will take the first step towards introducing deep
learning to the area of HPC performance evaluation, opening
the door for others. Convolutional neural networks will be
fed images of timeline views of HPC applications and will
identify the intrinsic behavior of the application and return
some principal performance metrics.
The results show that deep learning techniques indeed can
be utilized for evaluating the performance of parallel applications,
with the main limitation for its success being the sizes
of the data sets available. Furthermore a number of exciting
directions for taking the next step utilizing deep learning
techniques with performance evaluation are suggested. |