High Performance Optimization at the Door of the Exascale

TL;DR

This paper discusses the challenges and considerations in optimizing computational performance at the exascale level, focusing on hardware complexity, heterogeneity, memory hierarchy, and their impact on large-scale scientific and AI optimization problems.

Contribution

It provides an analysis of the key hardware and software challenges in achieving high performance on exascale supercomputers, emphasizing optimization techniques for complex, large-scale problems.

Findings

Hardware heterogeneity complicates optimization strategies.

Memory hierarchy impacts data access efficiency.

Irregular control flow and complex memory access hinder large-scale optimization.

Abstract

quest for processing speed potential. In fact, we always get a fraction of the technically available computing power (so-called {\em theoretical peak}), and the gap is likely to go hand-to-hand with the hardware complexity of the target system. Among the key aspects of this complexity, we have: the {\em heterogeneity} of the computing units, the {\em memory hierarchy and partitioning} including the non-uniform memory access (NUMA) configuration, and the {\em interconnect} for data exchanges among the computing nodes. Scientific investigations and cutting-edge technical activities should ideally scale-up with respect to sustained performance. The special case of quantitative approaches for solving (large-scale) problems deserves a special focus. Indeed, most of common real-life problems, even when considering the artificial intelligence paradigm, rely on optimization techniques for the main kernels of algorithmic solutions. Mathematical programming and pure combinatorial methods are not easy to implement efficiently on large-scale supercomputers because of {\em irregular control flow}, {\em complex memory access patterns}, {\em heterogeneous kernels}, {\em numerical issues}, to name a few. We describe and examine our thoughts from the standpoint of large-scale supercomputers.