PromiseTune: Unveiling Causally Promising and Explainable Configuration Tuning DOI: XXXXXXX.XXXXXXX Conference: the 48th IEEE/ACM International Conference on Software Engineering; April 12–18, 2026; Rio de Janeiro, Brazil ISBN: 978-1-4503-XXXX-X/2018/06 CCS: Software and its engineering Search-based

Pengzhou Chen, Tao Chen · 2025

Read PromiseTune: Unveiling Causally Promising and Explainable Configuration Tuning DOI: XXXXXXX.XXXXXXX Conference: the 48th IEEE/ACM International Conference on Software Engineering; April 12–18, 2026; Rio de Janeiro, Brazil ISBN: 978-1-4503-XXXX-X/2018/06 CCS: Software and its engineering Search-based by Pengzhou Chen, Tao Chen (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2507-05995.txt.

Introduction

Pengzhou Chen and Tao Chen present PromiseTune, a tool for tuning software settings, in this 2025 paper. Modern software has so many settings that finding a fast mix is hard, and testing each mix is costly. Tuners must choose between trying unknown mixes and refining mixes already known to be good.

PromiseTune learns rules that mark out promising regions of the setting space. It then filters those rules with causal reasoning, keeping only the rules that reflect real effects. Tuning then focuses inside the kept regions. Against eleven rival tuners on twelve systems, PromiseTune ranked far ahead overall. The kept rules also explain, in plain spatial terms, which parts of the setting space work well and why.

Catalog record

CREATORPengzhou Chen, Tao Chen
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE74 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2507-05995

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

PromiseTune: Unveiling Causally Promising and Explainable Configuration Tuning DOI: XXXXXXX.XXXXXXX Conference: the 48th IEEE/ACM International Conference on Software Engineering; April 12–18, 2026; Rio de Janeiro, Brazil ISBN: 978-1-4503-XXXX-X/2018/06 CCS: Software and its engineering Search-based software engineering CCS: Software and its engineering Software configuration management and version control systems CCS: Software and its engineering Software performance Pengzhou Chen, Tao Chen Abstract . The high configurability of modern software systems has made configuration tuning a crucial step for assuring system performance, e.g., latency or throughput. However, given the expensive measurements, large configuration space, and rugged configuration landscape, existing tuners suffer ineffectiveness due to the difficult balance of budget utilization between exploring uncertain regions (for escaping from local optima) and exploiting guidance of known good configurations (for fast convergence). The root cause is that we lack knowledge of where the promising regions lay, which also causes challenges in the explainability of the results. In this paper, we propose PromiseTune that tunes the configuration guided by causally purified rules. PromiseTune is unique in the sense that we learn rules, which reflect certain regions in the configuration landscape, and purify them with causal inference. The remaining rules serve as approximated reflections of the promising regions, bounding the tuning to emphasize these places in the landscape. This, as we demonstrate, can effectively mitigate the impact of the exploration and exploitation trade-off. Those purified regions can then be paired with the measured configurations to provide spatial explainability at the landscape level. Compared with 11 state-of-the-art tuners on 12 systems and varying budgets, we show that PromiseTune performs significantly better than the others with 42 % 42\% superior rank to the overall second best while providing richer information to explain the hidden system characteristics. 1. Introduction Software systems are becoming increasingly configurable, providing great flexibility to software users. However, this also incurs the difficulty of how to tune the configuration since it can profoundly impact system performance, e.g., latency and throughput ( Chen and Li 2023a ) . For example, it has been shown that for Storm , the default configuration can cause the system 480\times slower than the optimal one ( Jamshidi and Casale 2016 ) . Configuration tuning is therefore an important task in software engineering, as what have been reported in the literature ( Sayagh et al. 2020 ) . Yet, tuning complex systems is challenging, because: (a) (b) Figure 1 . Example of Redis system. (a) is the projected configuration landscape; (b) is the tuning trajectories of two tuners. • The number of possible configurations can be huge. For example, for the system 7z , 14 options have already led to more than a million configurations. • Configuration landscape is highly rugged/sparse ( Chen et al. 2025 ; Gong and Chen 2025 ) , meaning that there can be different local optima that might “trap” the tuning (see Figure 1 a). This makes sense, because if an option is to change the cache strategy, then it would significantly impact the performance. However, in the tuning, it is merely represented as a single-digit change. • The measurement can be extremely expensive ( Zhu et al. 2017 ; Chen et al. 2019 ; Chen and Li 2024 ; Chen and Li 2021 ; Ma et al. 2025 ) . For example, it takes more than 1,536 hours to sample the configurations of 11 options for x264 ( Valov et al. 2017 ) . Therefore, tens or hundreds of measurements are common budgets ( Nair et al. 2020 ; Chen et al. 2021 ) . As such, exhaustively profiling configurable system is unrealistic, thus in the past decade the research community has proposed various tuners based on heuristics ( Chen and Li 2023b ; Ye et al. 2025 ; Chen et al. 2019 ; Chen and Li 2024 ; Chen and Li 2021 ; Behzad et al. 2013 ; Shahbazian et al. 2020 ; Nair et al. 2020 ; Zhu and Hao 2023 ) , which are less sensitivity to the size of configuration space. However, those tuners often need to handle a difficult balance of how to spend the budget: either exploring uncertain regions (for escaping from local optima) or exploiting the known good configurations to guide (for fast convergence) ( Chen et al. 2024 ) . The former refers to exploration, meaning that more budget would be consumed for randomly jumping out from local optima under uncertainty, but there is no guarantee that the budget used would bring benefits; while the latter, which focuses on exploitation, uses more budget to search around the good configurations found so far, but it might easily lead to premature convergence at local optima. Because of the above, existing tuners can still struggle to tune certain systems. The fact that those tuners are mostly black-box further exacerbates this issue, as there is no explainability provided on the configuration landscape. Figure 1 b shows an example: we see that both the model-based tuner SMAC ( Hutter et al. 2011a ) and Random Search struggle, but due to completely opposed causes: SMAC adopts a greedy local search heuristic in the model space, hence it is highly efficient in using the budget to guide the tuning based on good configurations found, but can easily be trapped at local optima with premature convergence. In contrast, Random Search is naturally resilient to local optima, but it lacks strong guidance to efficiently utilize the budget for converging. In this paper, we take a different perspective on the above limitation and challenges: drawing on the observation that, in general, most of the good configurations tend to be more condensed to certain promising region(s) in the configuration space ( Nair et al. 2020 ; Chen et al. 2025 ; Chen et al. 2024 ) , we hypothesize that lacking the knowledge of those promising regions can be the root cause of the above ineffective tuning, complicating the issue of balancing budget for exploration and exploitation. To that end, we present PromiseTune , a tuner guided by causally-verified promising regions with explainability. The key idea is that we learn rules that bound the configuration landscape as the representation of regions and exploit causal inference to causally purify the rules that approximately reflect the promising regions. These rules, which can be iteratively updated and are self-explainable, would then guide a model-based Bayesian optimizer, mitigating the impact of exploration and exploitation trade-offs. What makes PromiseTune unique is that, unlike existing work where causality has been used to analyze configuration options ( Iqbal et al. 2022 ; Hossen et al. 2024 ) , we use it to purify the regions in the configuration space, as represented by rules, hence providing finer-grained control over the tuning. The purified rules, after further filtering using all measured configurations by the end of tuning, can be used to better explain the behaviors of the configurable system at a fine-grained landscape level. In a nutshell, our contributions are: • We extract the paths learned by a Random Forest—which is predominantly used in the configuration tuning and handles sparsity well ( Chen et al. 2021 ; Hutter et al. 2011a ) —as the rules and featurize them with the measured configurations, making them causally analyzable. • Rules are

Read the complete text