ABSTRACT: Many scientific applications are computationally intensive, data
parallel, irregular and contain loops with various characteristics and
large number of iterations. Over the years, to address multiple issues of
their performance and scalability in heterogeneous and unpredictable
changing environments, a number of dynamic loop scheduling (DLS) techniques
based on probabilistic analyses have been developed.
Recently, robustness metrics corresponding to each of the earlier
developed DLS techniques have been derived to ensure a reliable execution
for a specifically required guaranteed level of performance.
This research work is concerned with the development of reinforcement
learning (RL) agents, capable to make use of robustness metrics in the
optimal selection of the DLS technique that ensures the required level of
reliability and guaranteed performance for applications that include a
number of loops with different characteristics and also a large number of
time-steps. The ongoing work will be presented in the poster session.