Talk Title: Performance Prediction on Heterogeneous Architectures: Challenges and Insights
Presenter: Silvio Stanzani (Universidade Estadual Paulista)
Abstract:
Predicting the performance of parallel applications in different computational architectures is essential to improve clusters’ utilization, providing insights about the design of better job schedulers. In order for these schedulers to be able to make runtime decisions, two main constraints must be considered: 1) the application is provided without the source code and 2) the prediction must be performed within a small time limit. In order to overcome these challenges we use profilers, which measure various aspects of hardware usage during application execution. we will discuss the challenges and practical issues related to the implementation of an strategy that implements performance prediction based on information provided by profiler and on the Roofline model to characterize the efficiency that an application is using different computational architectures.
Biography: Silvio Stanzani works as Research Associate at UNESP Center for Scientific Computing since 2015 on projects related to Machine Learning and High Performance Computing, with focus on performance optimization and on teaching of parallel programming and data analytics. He holds a Ph.D. from Escola Politécnica da Universidade de São Paulo (EPUSP), and a Master from Universidade Católica de Santos (Unisantos). He works on software development focused on parallel programming and distributed computing since 2005.