Revealing Domain-Spatiality Patterns for Configuration Tuning: Domain Knowledge Meets Fitness Landscapes

Yulong Ye; Hongyuan Liang; Chao Jiang; Miqing Li; Tao Chen

cs.SEMar 20, 2026

Revealing Domain-Spatiality Patterns for Configuration Tuning: Domain Knowledge Meets Fitness Landscapes

Authors: Yulong Ye, Hongyuan Liang, Chao Jiang, Miqing Li, Tao Chen

Organizations: IDEAS Lab, University of Birmingham, United Kingdom · University of Electronic Science and Technology of China, China · University of Birmingham, United Kingdom

Abstract

Configuration tuning for better performance is crucial in quality assurance. Yet, there has long been a mystery on tuners' effectiveness, due to the black-box nature of configurable systems. Prior efforts predominantly adopt static domain analysis (e.g., static taint analysis), which often lacks generalizability, or dynamic data analysis (e.g., benchmarking performance analysis), limiting explainability. In this work, we embrace Fitness Landscape Analysis (FLA) as a bridge between domain knowledge and difficulty of the tuning. We propose Domland, a two-pronged methodology that synergizes the spatial information obtained from FLA and domain-driven analysis to systematically capture the hidden characteristics of configuration tuning cases, explaining how and why a tuner might succeed or fail. This helps to better interpret and contextualize the behavior of tuners and inform tuner design. To evaluate Domland, we conduct a case study of nine software systems and 93 workloads, from which we reveal several key findings: (1) configuration landscapes are inherently system-specific, with no single domain factor (e.g., system area, programming language, or resource intensity) consistently shaping their structure; (2) the core options (e.g., pic-struct of x264), which control the main functional flows, exert a stronger influence on landscape ruggedness (i.e. the difficulty of tuning) compared to resource options (e.g., cpu-independent of x264); (3) Workload effects on landscape structure are not uniformly tied to type or scale. Both contribute to landscape variations, but their impact is system-dependent.

Figures & tables

Figure 1. The regret (i.e., gap to the optimal performance) using Random Search (RS) and Hill Climbing (HC) to tune Batik and H2 . The top shows their landscapes processed by MDS ( Cox and Cox, 2008 ) .

Figure 2. Illustration of configuration landscape and performance distribution.

Figure 3. Overview of the Domland methodology.

Landscape	Metric	Interpretation	Implications for tuner design/selection
Global structure	FDC ( $\varrho$ ) ( Jones et al., 1995 )	Guided: $\varrho\geq 0.15$	A high correlation indicates that better configurations lie closer to the optimum, revealing predictable structural characteristics that favour exploitation-driven strategies (e.g., hill climbing ( Li et al., 2014 ; Xi et al., 2004 ) ). In contrast, weak or negative correlations illustrate that the landscape provides little or misleading structural signal, requiring exploration-driven methods (e.g., SWAY ( Chen et al., 2018a ) , GA ( Olaechea et al., 2014 ; Bäck and Schwefel, 1993 ) ).
		Irregular: $-0.15\textless\varrho\textless 0.15$
		Deceptive: $\varrho\leq-0.15$
	Basin of Attraction ( Malan, 2021 )	$>0.21$ : easy to reach optimum	A large global basin implies that even simple local search can easily reach the optimum. A small basin, by contrast, calls for exploration-enhancing mechanisms such as restarts or diversity injection to avoid premature convergence (e.g., ParamILS ( Hutter et al., 2009 ) ).
		$\textless 0.01$ : hard to reach optimum
Local structure	Number ( $\ell_{p}$ )	Large: high multimodality	A large number of local optima increases the risk of premature convergence, favoring diversity-preserving methods (e.g., multi-objectivization ( Chen et al., 2024 ; Chen and Li, 2021 ) ). Fewer optima allow effective use of deterministic local search (e.g., hill climbing ( Li et al., 2014 ; Xi et al., 2004 ) ).

Table 1. Summary of landscape metrics and their implications for configuration tuning.

Type		Descriptions	Exampled Option
Functional	Core ( $F_{1}$ )	Governs the system’s core logic, directly impacting execution behavior.	MVSTORE in H2
Functional	Utility ( $F_{2}$ )	Provides auxiliary functionalities applicable across multiple modules.	anyScriptOrigin in Batik
Resource	CPU ( $R_{1}$ )	Controls CPU usage, parallel processing, or multi-threading.	threads in Dconvert
	Storage ( $R_{2}$ )	Manages persistent data storage.	VBRRange in Jump3r
	Memory ( $R_{3}$ )	Handles cache, memory pools, and buffer.	blocksize in Kanzi
	Queue ( $R_{4}$ )	Regulates scheduling, timeouts, and delays.	FlushTimeout in Xz

Table 2. Domain features of option for configurable systems.

Panel of Figure 4 — Figure 4. Prompt template. ab denotes the content that will be substituted by actual data.

System	Language	Area	Performance	Version	Resource Intensity	LOC	#O	#C	#W
Jump3r ( , 2023 )	Java	Audio Encoder	Runtime	1.0.4	CPU	29685	13	4196	6
Kanzi ( , 2023 )	Java	File Compressor	Runtime	1.9	I/O; Memory	28614	18	4112	9
Dconvert ( , 2023 )	Java	Image Scaling	Runtime	1.0.0- $\alpha 7$	I/O; Memory	6888	10	6764	12
H2 ( , 2023 )	Java	Database	Throughput	1.4.200	I/O; Memory; Storage	333539	16	1954	8
Batik ( , 2023 )	Java	SVG Rasterizer	Runtime	1.14	CPU	360924	8	1919	11
Xz ( , 2023 )	C/C++	File Compressor	Runtime	5.2.0	CPU	43130	12	1999	13

Table 3. Subject system characteristics. The workload details can be accessed via https://tinyurl.com/546svkb2 .

System		7z				LLVM
Method	Data Scale	$\ell_{p}$	$\ell_{q}$	Basin of Attraction	$d$	$\ell_{p}$	$\ell_{q}$	Basin of Attraction	$d$
	Large (50%)	0.00758	0.85737	0.02092	1	0.01230	0.32979	0.08344	1
	Medium (25%)	0.03963	0.54504	0.00804	1	0.12177	0.15062	0.02307	1
Fixed	Small (5%)	0.43036	0.11389	0.00000	1	0.66575	0.02754	0.00000	1
	Large (50%)	0.00758	0.85737	0.02092	1	0.00009	0.92532	0.68311	2
	Medium (25%)	0.03963	0.54504	0.00804	1	0.00018	0.90927	0.82104	2

Table 4. Comparison of fitness landscape metrics computed under different neighborhood definitions and sampling granularities, where results from the full-scale dataset serve as the ground truth.

\ell_{p}

denotes the proportion of local optima;

\ell_{q}

represents the quality relative to the global optimum;

d

is neighborhood radius. “Fixed” uses a constant neighborhood radius, i.e.,

d=1

, whereas “Adaptive” applies the proposed sparsity-aware radius.

Figure 5. Comparison of FDC (

\varrho

) under different sampling densities. The red dashed line indicates the FDC value computed from the full-scale dataset (ground truth).

Lrzip

Dconvert

Batik

Kanzi

X264

Jump3r

Option

Type

RSD (%)

Type

RSD (%)

Type

RSD (%)

Type

RSD (%)

Type

RSD (%)

Type

RSD (%)

Type

RSD (%)

Type

RSD (%)

Type

RSD (%)

F_{1}

107.254

R_{3}

2.118

F_{2}

1.333

F_{1}

7.422

F_{2}

0.14

F_{1}

1.136

F_{1}

1.162

R_{1}

6.703

F_{1}

1.695

F_{2}

210.332

F_{1}

7.29

F_{2}

3.881

F_{1}

1.105

F_{2}

0.376

F_{1}

121.468

F_{1}

6.467

R_{1}

5.301

F_{1}

✗

F_{2}

83.073

F_{1}

19.927

F_{1}

0.276

F_{2}

0.2

F_{1}

0.705

F_{1}

0.081

F_{1}

11.448

R_{1}

5.947

F_{1}

1.412

R_{1}

499.477

R_{3}

20.944

F_{1}

1.182

F_{1}

0.099

F_{2}

1.224

F_{1}

164.22

F_{1}

8.785

R_{1}

0.133

F_{1}

✗

F_{1}

306.08

F_{1}

2.042

F_{1}

2.047

F_{1}

0.615

F_{2}

0.476

F_{1}

32.901

F_{1}

13.159

R_{1}

0.152

F_{1}

9.008

Table 5. The relative standard deviation (RSD) of autocorrelation across configuration options in nine software systems. For each case, blue cells indicate options with significant influence on landscape ruggedness (RSD

\geq

5%); or red cells otherwise. The “Type” column denotes the functional-related or resource-related roles of options as categorized in Table 2 . The ✗ markers indicate that there are insufficient data samples for analysis.

Workload	Lrzip	Xz	Z3	Dconvert	Batik	Kanzi	X264	H2	Jump3r
W1	0.025	0.381	0.488	0.461	0.319	0.207	-0.083	0.637	0.439
W2	0.186	0.460	0.535	0.524	0.280	0.328	-0.015	0.627	0.067
W3	-0.037	-0.198	0.545	0.468	0.520	0.123	-0.014	0.641	0.363
W4	-0.078	-0.099	0.492	0.511	0.550	0.181	0.065	0.257	0.267
W5	-0.050	0.566	0.098	0.523	0.360	0.123	0.031	0.637	0.443
W6	0.019	0.612	0.018	0.524	0.438	0.192	0.155	0.653	0.117

Table 6. The FDC (

\varrho

) across nine systems with different workloads ( blue cells :

\varrho\geq

0.15; red cells :

\mathcal{-}0.15

\varrho

<0.15; gold cells :

\varrho\leq\mathcal{-}

0.15).

Wrokload	Lrzip	Xz	Z3	Dconvert	Batik	Kanzi	X264	H2	Jump3r
W1	0.037	0.017	0.643	0.079	0.162	0.35	0.052	0.079	0.037
W2	0.565	0.449	0.735	0.107	0.11	0.239	0.004	0.06	0.081
W3	0.063	0.04	0.753	0.051	0.077	0.081	0.004	0.033	0.074
W4	0.052	0.121	0.142	0.107	0.09	0.301	0.013	0.045	0.109
W5	0.131	0.116	0.542	0.146	0.176	0.331	0.02	0.061	0.101
W6	0.304	0.102	0.642	0.085	0.109	0.11	0.003	0.079	0.032

Table 7. The proportion of configurations in the dataset that falls within the global optimum’s basin of attraction. For each cases, the blue cells indicate basin size of the global optimum is larger than 0.21; or red cells otherwise.

	Lrzip		Xz		Z3		Dconvert		Batik		Kanzi		X264		H2		Jump3r
Workload	$\ell_{p}$	$\ell_{q}$	$\ell_{p}$	$\ell_{q}$	$\ell_{p}$	$\ell_{q}$	$\ell_{p}$	$\ell_{q}$	$\ell_{p}$	$\ell_{q}$	$\ell_{p}$	$\ell_{q}$	$\ell_{p}$	$\ell_{q}$	$\ell_{p}$	$\ell_{q}$	$\ell_{p}$	$\ell_{q}$
W1	0.120	0.972	0.025	0.767	0.015	0.628	0.006	0.787	0.017	0.754	0.086	-0.333	0.180	0.096	0.026	0.244	0.041	0.376
W2	0.194	0.992	0.203	0.592	0.125	0.784	0.005	0.753	0.009	0.906	0.105	-0.234	0.164	0.072	0.025	0.071	0.042	0.534
W3	0.099	0.987	0.025	0.749	0.024	0.967	0.026	0.569	0.010	0.788	0.147	0.421	0.176	0.098	0.024	0.232	0.044	0.510
W4	0.105	0.945	0.021	0.818	0.016	0.966	0.005	0.739	0.011	0.890	0.084	-0.338	0.173	0.075	0.028	0.316	0.048	0.450
W5	0.099	0.830	0.025	0.633	0.012	0.992	0.008	0.718	0.011	0.879	0.086	-0.247	0.176	0.092	0.030	0.101	0.042	0.396

Table 8. The proportion and quality of local optima across workloads for each system. The blue cells , red cells , and gold cells indicate local optima with high (

\ell_{q}\geq 0.67

), medium (0.33

\leq\ell_{q}\leq

0.67), and low quality (

\ell_{q}\leq 0.33

Figure 6. The Spearman correlation between local optima superiority and basin size across workloads per system.

Workload	Lrzip	Xz	Z3	Dconvert	Batik	Kanzi	X264	H2	Jump3r
W1	-0.012	0.630	0.659	0.781	0.688	0.339	0.254	0.779	0.411
W2	0.167	0.196	0.697	0.779	0.734	0.365	0.263	0.828	0.320
W3	0.017	0.653	0.604	0.441	0.796	0.329	0.258	0.770	0.400
W4	-0.034	0.491	0.729	0.772	0.825	0.316	0.244	0.129	0.420
W5	-0.007	0.626	0.605	0.754	0.736	0.316	0.211	0.807	0.442
W6	0.026	0.597	0.617	0.591	0.725	0.336	0.282	0.803	0.438

Table 9. The autocorrelation across different workloads for nine software configurable systems. The blue cells , red cells , and gold cells indicate smooth landscapes (

r\geq 0.5

), moderately rugged landscapes (

0.2\textless r\textless 0.5

), and highly rugged landscapes (

r\leq 0.2

), respectively.

Dconvert		H2
Workload (Type-Scale)	Autocorrelation	Workload (Type-Scale)	Autocorrelation
jpeg-large	0.781	smallbank-1	0.779
jpeg-medium	0.779	smallbank-10	0.828
jpeg-small	0.441	tpcc-2	0.770
png-large	0.772	tpcc-8	0.129
png-medium	0.754	voter-16	0.807

Table 10. Autocorrelation of workloads in Dconvert and H2 .

Figure 7. The regret trajectories of Vanilla HC and Priority HC, aggregated across all workloads for each system.

Figure 8. The regret trajectories of Vanilla GA and Priority GA, aggregated across all workloads for each system.

Figure 9. The regret trajectories using GA, RS, HC, and TPE to tune nine systems.

Panel of Figure 10 — Figure 10. Illustration of the configuration landscape and local optima for workload W1 of Lrzip . (a) Normalized negative runtime landscape (higher values indicate better performance). (b) Spatial distribution of identified local optima and their corresponding performance values.

		Lrzip	Xz	Z3	Dconvert	Batik	Kanzi	X264	H2	Jump3r
Workload	Tuner	Mean (Std)	Mean (Std)	Mean (Std)	Mean (Std)	Mean (Std)	Mean (Std)	Mean (Std)	Mean (Std)	Mean (Std)
	Transfer GA	3.131 (0.014)	3.773 (0.955)	5.860 (0.016)	1.797 (0.009)	0.903 (0.009)	1.129 (1.529)	0.929 (0.258)	26336.284 (1348.381)	2.376 (0.478)
W1	Restarted GA	3.133 (0.017)	4.732 (1.299)	5.907 (0.226)	1.860 (0.092)	0.924 (0.019)	1.832 (1.368)	1.097 (0.348)	25721.608 (816.896)	2.768 (0.658)
	Transfer GA	0.030 (0.000)	0.011 (0.003)	2.333 (0.656)	1.078 (0.040)	1.334 (0.006)	0.133 (0.048)	3.778 (1.110)	18266.502 (1838.219)	0.753 (0.131)
W2	Restarted GA	0.030 (0.000)	0.014 (0.005)	1.761 (0.004)	1.110 (0.040)	1.360 (0.021)	0.160 (0.044)	4.005 (1.024)	18378.785 (822.054)	0.934 (0.272)
	Transfer GA	3.301 (0.003)	4.014 (1.188)	0.402 (0.695)	0.371 (0.003)	4.177 (0.020)	0.427 (0.311)	1.322 (0.300)	932.575 (51.553)	1.270 (0.357)

Table 11. The Mean and Standard deviation (Std) of performance objectives between Transfer GA and Restarted GA for all cases over 30 runs; The blue cells , red cells , and gold cells respectively indicate Transfer GA performing significantly better than, similarly to, or worse than Restarted GA.

Fitness Landscape Analysis

See all 12 papers

Less Is More: Tuning Configurable Systems with Imperfect Fidelity
Aug 1, 2026Yulong Ye, Miqing Li, Tao ChenMulti-Fidelity Optimization
IDEAS Lab · University of Birmingham · Birmingham, United Kingdom
When to Use Which? Benchmarking Optimisers for Configurable Systems under Varying Budgets
Jul 17, 2026Chao Jiang, Yulong Ye, Tao Chen +1Black-Box Optimization Software Engineering Benchmarks
University of Birmingham, United Kingdom · IDEAS Lab, University of Birmingham, United Kingdom
CATune: Structural Constraint-Aware Bayesian Optimization for DBMS Configuration Tuning
Oct 7, 2026Fangping Lan, Qi Zhang, Eduard DragutDatabase Systems Bayesian Optimization
Temple University Philadelphia, PA, USA

Aug 1, 2026J/K move · Enter open · S save

Less Is More: Tuning Configurable Systems with Imperfect Fidelity

Yulong Ye, Miqing Li, Tao Chen

IDEAS Lab · University of Birmingham · Birmingham, United Kingdom

Multi-Fidelity Optimization

Open paper AskPDF arXiv

Configuration tuning is essential for optimizing the performance of highly configurable systems, e.g., throughput or runtime, under a given environment. Yet, this is a challenging process as there can be many options to tune, and configuration measurement is often highly expensive. In this paper, we demonstrate the phenomenon of ``less can be more'': system configuration tuning can be greatly improved with much superior budget utilization by partially tuning under the imperfect-fidelity---an environment that is similar, but cheaper to measure, compared with the concerned perfect-fidelity of environment under which the system should be tuned. We codify a conceptual framework of fidelity for configurable systems, drawing on which allows us to propose MFTune, a tuner that proactively explores in the space of

>10^4

possible imperfect-fidelity settings to approximate a useful one, which strikes for the wideness of tuning. This creates high-quality seeds for the perfect-fidelity, which in turn ensures the tuning depth. Experiment results against

10

state-of-the-art tuners, obtained from running diverse real-world systems for

19

months

24 \times 7

, show that MFTune performs considerably better on

83.33

% cases with up to

19.34\%

improvement while achieving hours of budget saving in general.

Abstract

Fitness Landscape Analysis

Explore similar work

Less Is More: Tuning Configurable Systems with Imperfect Fidelity

When to Use Which? Benchmarking Optimisers for Configurable Systems under Varying Budgets

CATune: Structural Constraint-Aware Bayesian Optimization for DBMS Configuration Tuning