CONQURE: A Co-Execution Environment for Quantum and Classical Resources

TL;DR

CONQURE introduces an open-source co-execution environment that seamlessly integrates quantum and classical computing resources, enabling efficient offloading and scheduling of quantum kernels within HPC and ML workflows.

Contribution

It presents a modular scheduling framework and API for quantum offloading, with demonstrated low overhead and performance improvements in quantum algorithm execution.

Findings

API overhead averages 12.7ms

Enables parallel VQE runs with 3.1X speedup

Demonstrated on an ion-trap quantum device

Abstract

Cutting edge classical computing today relies on a combination of CPU-based computing with a strong reliance on accelerators. In particular, high-performance computing (HPC) and machine learning (ML) rely heavily on acceleration via GPUs for numerical kernels. In the future, acceleration via quantum devices may complement GPUs for kernels where algorithms provide quantum advantage, i.e., significant speedups over classical algorithms. Computing with quantum kernels mapped onto quantum processing units (QPUs) requires seamless integration into HPC and ML. However, quantum offloading onto HPC/cloud lacks open-source software infrastructure. For classical algorithms, parallelization standards, such as OpenMP, MPI, or CUDA exist. In contrast, a lack of quantum abstractions currently limits the adoption of quantum acceleration in practical applications creating a gap between quantum algorithm development and practical HPC integration. Such integration needs to extend to efficient quantum offloading of kernels, which further requires scheduling of quantum resources, control of QPU kernel execution, tracking of QPU results, providing results to classical calling contexts and coordination with HPC scheduling. This work proposes CONQURE, a co-execution environment for quantum and classical resources. CONQURE is a fully open-source cloud queue framework that presents a novel modular scheduling framework allowing users to offload OpenMP quantum kernels to QPUs as quantum circuits, to relay results back to calling contexts in classical computing, and to schedule quantum resources via our CONQURE API. We show our API has a low overhead averaging 12.7ms in our tests, and we demonstrate functionality on an ion-trap device. Our OpenMP extension enables the parallelization of VQE runs with a 3.1X reduction in runtime.