Comparing performance of variational quantum algorithm simulations on HPC systems

TL;DR

This paper presents a toolchain that standardizes problem definitions for variational quantum algorithms, enabling consistent comparisons across different simulators and HPC environments, while highlighting scalability challenges.

Contribution

The authors develop a generic problem description and parser tool for variational quantum algorithms, facilitating cross-simulator comparisons and performance analysis on HPC systems.

Findings

The tool successfully translates problem definitions between simulators.

Variational algorithms face scalability limitations due to long runtimes and memory constraints.

Job arrays partially mitigate parallelism limitations.

Abstract

Variational quantum algorithms are of special importance in the research on quantum computing applications because of their applicability to current Noisy Intermediate-Scale Quantum (NISQ) devices. The main building blocks of these algorithms (among them, the definition of the Hamiltonian and of the ansatz, the optimizer) define a relatively large parameter space, making the comparison of results and performance between different approaches and software simulators cumbersome and prone to errors. In this paper, we employ a generic description of the problem, in terms of both Hamiltonian and ansatz, to port a problem definition consistently among different simulators. Three use cases of relevance for current quantum hardware (ground state calculation for the Hydrogen molecule, MaxCut, Travelling Salesman Problem) have been run on a set of HPC systems and software simulators to study the dependence of performance on the runtime environment, the scalability of the simulation codes and the mutual agreement of the physical results, respectively. The results show that our toolchain can successfully translate a problem definition between different simulators. On the other hand, variational algorithms are limited in their scaling by the long runtimes with respect to their memory footprint, so they expose limited parallelism to computation. This shortcoming is partially mitigated by using techniques like job arrays. The potential of the parser tool for exploring HPC performance and comparisons of results of variational algorithm simulations is highlighted.