Performance Optimization of ML and HPC Applications on Heterogeneous Systems

Zhen Xie, University of California, Merced
Supercomputer showdown

Abstract: Heterogeneous systems have gained popularity due to the rapid growth in data and requirements from ML and HPC applications. GPU and TB-scale big memory become the basic modules to build massively parallel computing systems with large data throughput. In this talk, we discuss challenges and opportunities of optimizing ML and HPC applications on heterogeneous systems. And two case studies are used to demonstrate potential performance improvements. For ML case, we use the decision tree as an example. We rearrange tree nodes to enable efficient and coalesced memory accesses; and we also rearrange trees, such that trees with similar structures are grouped together in a balanced way. For HPC case, we use Molecular dynamics simulation as an example, and introduce a new method to improve the performance by leveraging the big memory. Two cases can outperform the state-of-the-art libraries with an average speedup of 3.8x and 7.6x. It also proves that heterogeneous system has great potential to accelerate ML and HPC applications.


Please use this link to attend the virtual seminar.

Meeting ID: 552 037 531 // Participant Code: 9606