arXiv subjects

Ben van Werkhoven

Publications and source records attributed to Ben van Werkhoven.

16 recordsLinked to original sources

arXiv2026

Constraint-aware Optimization in Auto-Tuning

Automatic performance tuning, or auto-tuning, is a key technique in high-performance computing, enabling applications to adapt to complex and evolving hardware architectures. A central challenge is the need to optimize over large discrete, constrained parameter spaces, where many candidate configurations are invalid due to hardware or software correctness constraints. Traditional evolutionary algorithms, such as Differential Evolution, Particle Swarm Optimization, and Genetic Algorithms, are not inherently constraint-aware and thus often waste computational resources evaluating invalid solutions. In this work, we present and evaluate constraint-aware variants of four evolutionary algorithms for auto-tuning. Through extensive experiments on a representative benchmark suite, we show that constraint-aware optimization leads to faster convergence and improved performance over unconstrained methods. Furthermore, we demonstrate that our methods outperform the pyATF methods, a state-of-the-art framework for constraint-based auto-tuning. Our results demonstrate that incorporating constraint-awareness into the optimization process significantly enhances their applicability and effectiveness in real-world auto-tuning problems. Constraint-awareness improved algorithm efficiency by ~39 on average, correlated with search space sparsity. The algorithms developed in this study are publicly available as open-source contributions to the Kernel Tuner framework, facilitating future research and benefitting users.

cs.NE↗

arXiv2026

Automated Algorithm Design for Auto-Tuning Optimizers

Automatic performance tuning (auto-tuning) is essential for optimizing high-performance applications, where vast and irregular search spaces make manual exploration infeasible. While auto-tuners traditionally rely on classical approaches such as evolutionary, annealing, or surrogate-based optimizers, designing algorithms that efficiently find near-optimal configurations robustly across diverse tasks is challenging. We propose a new paradigm: using large language models (LLMs) to automatically generate optimization algorithms tailored to auto-tuning problems. We introduce a framework that prompts LLMs with problem descriptions and search space characteristics to synthesize, test, and iteratively refine specialized optimizers. These generated algorithms are evaluated on four real-world auto-tuning applications across six hardware platforms and compared against the state-of-the-art in two contemporary auto-tuning frameworks. The evaluation demonstrates that providing additional application- and search space-specific information in the generation stage results in an average performance improvement of 30.7% and 14.6%, respectively. In addition, our results show that LLM-generated optimizers can rival, and in various cases outperform, existing human-designed algorithms, with our best-performing generated optimization algorithms achieving an average 72.4% improvement over state-of-the-art optimizers for auto-tuning.

cs.LG↗

arXiv2025

Tuning the Tuner: Introducing Hyperparameter Optimization for Auto-Tuning

Automatic performance tuning (auto-tuning) is widely used to optimize performance-critical applications across many scientific domains by finding the best program variant among many choices. Efficient optimization algorithms are crucial for navigating the vast and complex search spaces in auto-tuning. As is well known in the context of machine learning and similar fields, hyperparameters critically shape optimization algorithm efficiency. Yet for auto-tuning frameworks, these hyperparameters are almost never tuned, and their potential performance impact has not been studied. We present a novel method for general hyperparameter tuning of optimization algorithms for auto-tuning, thus "tuning the tuner". In particular, we propose a robust statistical method for evaluating hyperparameter performance across search spaces, publish a FAIR data set and software for reproducibility, and present a simulation mode that replays previously recorded tuning data, lowering the costs of hyperparameter tuning by two orders of magnitude. We show that even limited hyperparameter tuning can improve auto-tuner performance by 94.8% on average, and establish that the hyperparameters themselves can be optimized efficiently with meta-strategies (with an average improvement of 204.7%), demonstrating the often overlooked hyperparameter tuning as a powerful technique for advancing auto-tuning research and practice.

cs.LG↗

arXiv2025

Efficient Construction of Large Search Spaces for Auto-Tuning

Automatic performance tuning, or auto-tuning, accelerates high-performance codes by exploring vast spaces of code variants. However, due to the large number of possible combinations and complex constraints, constructing these search spaces can be a major bottleneck. Real-world applications have been encountered where the search space construction takes minutes to hours or even days. Current state-of-the-art techniques for search space construction, such as chain-of-trees, lack a formal foundation and only perform adequately on a specific subset of search spaces. We show that search space construction for constraint-based auto-tuning can be reformulated as a Constraint Satisfaction Problem (CSP). Building on this insight with a CSP solver, we develop a runtime parser that translates user-defined constraint functions into solver-optimal expressions, optimize the solver to exploit common structures in auto-tuning constraints, and integrate these and other advances in open-source tools. These contributions substantially improve performance and accessibility while preserving flexibility. We evaluate our approach using a diverse set of benchmarks, demonstrating that our optimized solver reduces construction time by four orders of magnitude versus brute-force enumeration, three orders of magnitude versus an unoptimized CSP solver, and one to two orders of magnitude versus leading auto-tuning frameworks built on chain-of-trees. We thus eliminate a critical scalability barrier for auto-tuning and provide a drop-in solution that enables the exploration of previously unattainable problem scales in auto-tuning and related domains.

cs.DC↗

arXiv2025

PowerSensor3: A Fast and Accurate Open Source Power Measurement Tool

Power consumption is a major concern in data centers and HPC applications, with GPUs typically accounting for more than half of system power usage. While accurate power measurement tools are crucial for optimizing the energy efficiency of (GPU) applications, both built-in power sensors as well as state-of-the-art power meters often lack the accuracy and temporal granularity needed, or are impractical to use. Released as open hardware, firmware, and software, PowerSensor3 provides a cost-effective solution for evaluating energy efficiency, enabling advancements in sustainable computing. The toolkit consists of a baseboard with a variety of sensor modules accompanied by host libraries with C++ and Python bindings. PowerSensor3 enables real-time power measurements of SoC boards and PCIe cards, including GPUs, FPGAs, NICs, SSDs, and domain-specific AI and ML accelerators. Additionally, it provides significant improvements over previous tools, such as a robust and modular design, current sensors resistant to external interference, simplified calibration, and a sampling rate up to 20 kHz, which is essential to identify GPU behavior at high temporal granularity. This work describes the toolkit design, evaluates its performance characteristics, and shows several use cases (GPUs, NVIDIA Jetson AGX Orin, and SSD), demonstrating PowerSensor3's potential to significantly enhance energy efficiency in modern computing environments.

cs.PF↗

arXiv2024

Bringing Auto-tuning to HIP: Analysis of Tuning Impact and Difficulty on AMD and Nvidia GPUs

Many studies have focused on developing and improving auto-tuning algorithms for Nvidia Graphics Processing Units (GPUs), but the effectiveness and efficiency of these approaches on AMD devices have hardly been studied. This paper aims to address this gap by introducing an auto-tuner for AMD's HIP. We do so by extending Kernel Tuner, an open-source Python library for auto-tuning GPU programs. We analyze the performance impact and tuning difficulty for four highly-tunable benchmark kernels on four different GPUs: two from Nvidia and two from AMD. Our results demonstrate that auto-tuning has a significantly higher impact on performance on AMD compared to Nvidia (10x vs 2x). Additionally, we show that applications tuned for Nvidia do not perform optimally on AMD, underscoring the importance of auto-tuning specifically for AMD to achieve high performance on these GPUs.

cs.DC↗

arXiv2023

Kernel Launcher: C++ Library for Optimal-Performance Portable CUDA Applications

Graphic Processing Units (GPUs) have become ubiquitous in scientific computing. However, writing efficient GPU kernels can be challenging due to the need for careful code tuning. To automatically explore the kernel optimization space, several auto-tuning tools - like Kernel Tuner - have been proposed. Unfortunately, these existing auto-tuning tools often do not concern themselves with integration of tuning results back into applications, which puts a significant implementation and maintenance burden on application developers. In this work, we present Kernel Launcher: an easy-to-use C++ library that simplifies the creation of highly-tuned CUDA applications. With Kernel Launcher, programmers can capture kernel launches, tune the captured kernels for different setups, and integrate the tuning results back into applications using runtime compilation. To showcase the applicability of Kernel Launcher, we consider a real-world computational fluid dynamics code and tune its kernels for different GPUs, input domains, and precisions.

cs.DC↗

arXiv2023

Towards a Benchmarking Suite for Kernel Tuners

As computing system become more complex, it is becoming harder for programmers to keep their codes optimized as the hardware gets updated. Autotuners try to alleviate this by hiding as many architecture-based optimization details as possible from the user, so that the code can be used efficiently across different generations of systems. In this article we introduce a new benchmark suite for evaluating the performance of optimization algorithms used by modern autotuners targeting GPUs. The suite contains tunable GPU kernels that are representative of real-world applications, allowing for comparisons between optimization algorithms and the examination of code optimization, search space difficulty, and performance portability. Our framework facilitates easy integration of new autotuners and benchmarks by defining a shared problem interface. Our benchmark suite is evaluated based on five characteristics: convergence rate, local minima centrality, optimal speedup, Permutation Feature Importance (PFI), and performance portability. The results show that optimization parameters greatly impact performance and the need for global optimization. The importance of each parameter is consistent across GPU architectures, however, the specific values need to be optimized for each architecture. Our portability study highlights the crucial importance of autotuning each application for a specific target architecture. The results reveal that simply transferring the optimal configuration from one architecture to another can result in a performance ranging from 58.5% to 99.9% of the optimal performance, depending on the GPU architecture. This highlights the importance of autotuning in modern computing systems and the value of our benchmark suite in facilitating the study of optimization algorithms and their effectiveness in achieving optimal performance for specific target architectures.

cs.DC↗

arXiv2022

Going green: optimizing GPUs for energy efficiency through model-steered auto-tuning

Graphics Processing Units (GPUs) have revolutionized the computing landscape over the past decade. However, the growing energy demands of data centres and computing facilities equipped with GPUs come with significant capital and environmental costs. The energy consumption of GPU applications greatly depend on how well they are optimized. Auto-tuning is an effective and commonly applied technique of finding the optimal combination of algorithm, application, and hardware parameters to optimize performance of a GPU application. In this paper, we introduce new energy monitoring and optimization capabilities in Kernel Tuner, a generic auto-tuning tool for GPU applications. These capabilities enable us to investigate the difference between tuning for execution time and various approaches to improve energy efficiency, and investigate the differences in tuning difficulty. Additionally, our model for GPU power consumption greatly reduces the large tuning search space by providing clock frequencies for which a GPU is likely most energy efficient.

cs.DC↗

arXiv2022

Benchmarking optimization algorithms for auto-tuning GPU kernels

Recent years have witnessed phenomenal growth in the application, and capabilities of Graphical Processing Units (GPUs) due to their high parallel computation power at relatively low cost. However, writing a computationally efficient GPU program (kernel) is challenging, and generally only certain specific kernel configurations lead to significant increases in performance. Auto-tuning is the process of automatically optimizing software for highly-efficient execution on a target hardware platform. Auto-tuning is particularly useful for GPU programming, as a single kernel requires re-tuning after code changes, for different input data, and for different architectures. However, the discrete, and non-convex nature of the search space creates a challenging optimization problem. In this work, we investigate which algorithm produces the fastest kernels if the time-budget for the tuning task is varied. We conduct a survey by performing experiments on 26 different kernel spaces, from 9 different GPUs, for 16 different evolutionary black-box optimization algorithms. We then analyze these results and introduce a novel metric based on the PageRank centrality concept as a tool for gaining insight into the difficulty of the optimization problem. We demonstrate that our metric correlates strongly with observed tuning performance.

cs.DC↗

arXiv2022

Lightning: Scaling the GPU Programming Model Beyond a Single GPU

The GPU programming model is primarily aimed at the development of applications that run one GPU. However, this limits the scalability of GPU code to the capabilities of a single GPU in terms of compute power and memory capacity. To scale GPU applications further, a great engineering effort is typically required: work and data must be divided over multiple GPUs by hand, possibly in multiple nodes, and data must be manually spilled from GPU memory to higher-level memories. We present Lightning: a framework that follows the common GPU programming paradigm but enables scaling to large problems with ease. Lightning supports multi-GPU execution of GPU kernels, even across multiple nodes, and seamlessly spills data to higher-level memories (main memory and disk). Existing CUDA kernels can easily be adapted for use in Lightning, with data access annotations on these kernels allowing Lightning to infer their data requirements and the dependencies between subsequent kernel launches. Lightning efficiently distributes the work/data across GPUs and maximizes efficiency by overlapping scheduling, data movement, and kernel execution when possible. We present the design and implementation of Lightning, as well as experimental results on up to 32 GPUs for eight benchmarks and one real-world application. Evaluation shows excellent performance and scalability, such as a speedup of 57.2x over the CPU using Lighting with 16 GPUs over 4 nodes and 80 GB of data, far beyond the memory capacity of one GPU.

cs.DC↗

arXiv2021

Bayesian Optimization for auto-tuning GPU kernels

Finding optimal parameter configurations for tunable GPU kernels is a non-trivial exercise for large search spaces, even when automated. This poses an optimization task on a non-convex search space, using an expensive to evaluate function with unknown derivative. These characteristics make a good candidate for Bayesian Optimization, which has not been applied to this problem before. However, the application of Bayesian Optimization to this problem is challenging. We demonstrate how to deal with the rough, discrete, constrained search spaces, containing invalid configurations. We introduce a novel contextual variance exploration factor, as well as new acquisition functions with improved scalability, combined with an informed acquisition function selection mechanism. By comparing the performance of our Bayesian Optimization implementation on various test cases to the existing search strategies in Kernel Tuner, as well as other Bayesian Optimization implementations, we demonstrate that our search strategies generalize well and consistently outperform other search strategies by a wide margin.

cs.LG↗

arXiv2020

ESiWACE2 Services: RSE collaborations in Weather and Climate

We present the collaborative model of ESiWACE2 Services, where Research Software Engineers (RSEs) from the Netherlands eScience Center (NLeSC) and Atos offer their expertise to climate and earth system modeling groups across Europe. Within 6-month collaborative projects, the RSEs intend to provide guidance and advice regarding the performance, portability to new architectures, and scalability of selected applications. We present the four awarded projects as examples of this funding structure.

cs.SE↗

arXiv2020

Rocket: Efficient and Scalable All-Pairs Computations on Heterogeneous Platforms

All-pairs compute problems apply a user-defined function to each combination of two items of a given data set. Although these problems present an abundance of parallelism, data reuse must be exploited to achieve good performance. Several researchers considered this problem, either resorting to partial replication with static work distribution or dynamic scheduling with full replication. In contrast, we present a solution that relies on hierarchical multi-level software-based caches to maximize data reuse at each level in the distributed memory hierarchy, combined with a divide-and-conquer approach to exploit data locality, hierarchical work-stealing to dynamically balance the workload, and asynchronous processing to maximize resource utilization. We evaluate our solution using three real-world applications (from digital forensics, localization microscopy, and bioinformatics) on different platforms (from a desktop machine to a supercomputer). Results shows excellent efficiency and scalability when scaling to 96 GPUs, even obtaining super-linear speedups due to a distributed cache.

cs.DC↗

arXiv2020

Lessons learned in a decade of research software engineering GPU applications

After years of using Graphics Processing Units (GPUs) to accelerate scientific applications in fields as varied as tomography, computer vision, climate modeling, digital forensics, geospatial databases, particle physics, radio astronomy, and localization microscopy, we noticed a number of technical, socio-technical, and non-technical challenges that Research Software Engineers (RSEs) may run into. While some of these challenges, such as managing different programming languages within a project, or having to deal with different memory spaces, are common to all software projects involving GPUs, others are more typical of scientific software projects. Among these challenges we include changing resolutions or scales, maintaining an application over time and making it sustainable, and evaluating both the obtained results and the achieved performance. %In this paper, we present the challenges and lessons learned from research software engineering GPU applications.

cs.SE↗

arXiv2019

GPU acceleration of the SAGECal calibration package for the SKA

SAGECal has been designed to find the most accurate calibration solutions for low radio frequency imaging observations, with minimum artefacts due to incomplete sky models. SAGECAL is developed to handle extremely large datasets, e.g., when the number of frequency bands greatly exceeds the number of available nodes on a compute cluster. Accurate calibration solutions are derived at the expense of large computational loads, which require distributed computing and modern compute devices, such as GPUs, to decrease runtimes. In this work, we investigate if the GPU version of SAGECal scales well enough to meet the requirements for the Square Kilometre Array and we compare its performance with the CPU version.

astro-ph.IM↗