Towards a Dutch hybrid Quantum-HPC infrastructure
Quantum computers are devices that process information by taking advantage of the quantum-mechanical properties of their building blocks, the qubits. By doing so, they can harness work in a powerful and efficient way and perform certain operations with an exponential speed-up. There are many fields that could benefit from such a speed-up, for example, machine learning, financial modeling, logistic optimization, climate simulations, etc. Notably quantum computers are expected to excel at simulating quantum systems, like the ones present in chemistry and material science.
Hybrid algorithms are currently the only way to exploit NISQ devices and have been proposed for most applications envisioned for quantum computers. Hybrid algorithms require a hybrid approach to be executed. However, a hybrid approach is not exclusive to the execution of hybrid algorithms. Even for "pure" algorithms the pre- and post-processing is always done in a classical environment while the state preparation, unitary transformation, and measurements are done in the quantum environment. In other words, even after the NISQ era, it is expected that the execution of quantum algorithms (hybrid or not) will not only always require the support of classical computers but on the contrary, as quantum computers improve, the need for classical computing power to work closely and performantly with quantum computers will only grow. Quantum applications will always be executed by a hybrid quantum-classical workflow. Not only out of need but also because the existing application stacks are complex and the rewriting of high-level application software into the new quantum paradigm is ineffective and unlikely to happen. It is much more likely and advantageous to maximize the benefits of quantum computers by integrating their novelty into existing (classical) workflows.
Technical summary
The technical summary has been extracted from Schüsler et al. 2024, which was published in the context of this project.
To optimally execute hybrid quantum-classical applications we need to identify when and how classical resources play a role in the application. In this paper, we will refer to three main operational levels: application, task, and control (see Figure 1 in Reference 1).
- Application level: Most hybrid applications will be part of complex workflows executed in a main classical host. Only some tasks will be off-loaded to the quantum computer. The execution of the full application will require the orchestration of classical and quantum tasks. To ensure the quantum tasks do not slow down the high-level application, feedback latency of the order of seconds is desirable.
- Task level: A quantum task might require the on-loading of some sub-tasks to classical resources, for example, for quantum circuit cutting, circuit architecture search, classical optimization in variational algorithms, etc. To ensure the quantum task is not slowed down due to the on-loading of classical resources, feedback latency of the order of the quantum measurements is desirable. Although very variable and hardware-dependent, in general, a latency of ms should be expected. A lower feedback latency might be necessary to enable the execution of protocols that require e.g. mid-circuit measurements.
- Control level: The quantum control hardware executes gates and measurements in different manners depending on the quantum-chip technology. It generally also stores, reads, and interprets the received instructions in order to generate respective pulse sequences. To enable any useful control of a qubit, programmable control flow that operates on the timescale of nanoseconds is a requirement.
Each operational level manages computational resources differently. For the application level, assuming the main classical host is a High-Performance Computing (HPC) center, common examples of job managers are SLURM and OpenPBS. New developments on scheduling are done with tools like Flux, Hyperqueue, and QCGpilot due to their capability to run efficiently and easily on modern heterogeneous supercomputers.
For the task level, the resource manager needs to be able to work in the time scale of the QPU or lower. Although some existing job managers could be adapted to handle these timescales, the approach so far has been to develop job managers that take into account the specific needs of the quantum applications. The classical sub-tasks are executed in a classical runtime (see Figure 2). A classical runtime could be the main classical host (e.g. HPC Center or user laptop) or for example a small server in the proximity of the quantum computer. This last setup is often referred to as “hybrid runtime” and is for example used by Qiskit Runtime and Amazon Bracket Hybrid Jobs. The optimization of the communication between computational resources strongly depends on the hybrid infrastructure set-up (stand-alone, co-location, distributed).
At the control level, to perform logical operations, the hardware-specific instructions need to be very carefully timed by the classical control electronics. This is often achieved using a field programmable gate array (FPGA).
Due to the scarcity of quantum resources, an additional “job manager” is in some cases needed to orchestrate the quantum tasks arriving from different sources, for example, different users in an HPC center, different HPC centers, or in some cases individual users (see Figure 2 in Reference 1). Queue-based access might not be the most suitable when executing tasks that require a high-frequency coupling between the main classical host and the quantum computer. Instead, access to the quantum computer might have to be reserved for the duration of the task. Such a model has already been used by for example IBM Q.
ETP contribution
- Installed dependencies on ETP servers for the integration with Quantum Inspire.
- Investigated options for global installation such as creating an easy build.
- Collaborated with the project team to test job submissions and queries using tokens.
Outcomes
To cultivate competencies and skills within the team, understanding the requirements for integration, and identifying bottlenecks and areas in need of enhancement, with the ultimate goal of strengthening the team's knowledge base.
- To learn how to build hybrid infrastructures that integrate very different systems while recognizing the contrasting interests between Quantum Inspire and SURF. This awareness is crucial as it could result in diverse decision-making approaches, necessitating the discovery of the most effective approach.
To offer a collaborative hybrid Quantum–HPC computing platform.
References
Olaf Schüsler, Ariana Torres-Knoop, Jaap Dijkshoorn, Christiaan Hollemans, Bas van der Vlies, Richard Versluis. "Towards a Dutch hybrid quantum/HPC infrastructure," ArXiV, 2403.17649, 2024 (https://arxiv.org/abs/2403.17649).