MCTuner: Adaptive Database Knob Tuning

• MCTuner is an adaptive database tuning framework that integrates LLM-guided knob selection with hierarchical search over high-dimensional, mixed-type configuration spaces.
• It employs a two-stage pipeline where an LLM-guided Mixture-of-Experts compresses knob ranges followed by spatial decomposition with MCTS and localized Bayesian Optimization.
• Empirical results on PostgreSQL demonstrate up to 19.2% performance improvements and 1.4x faster configuration discovery compared to state-of-the-art tuners.

Searching arXiv for MCTuner and closely related tuning literature. arxiv_search(query="MCTuner database tuning", max_results=10) arxiv_search({"query":"MCTuner database tuning","max_results":10}) MCTuner is an adaptive database knob tuning framework for modern DBMSs that combines LLM-guided knob selection with hierarchical search over the configuration space. It is introduced as “Spatial Decomposition-Enhanced Database Tuning via LLM-Guided Exploration” and is designed for settings in which a DBMS exposes hundreds of continuous, integer, and categorical knobs whose joint effects determine performance, stability, and cost (Yan et al., 8 Sep 2025). The framework addresses two stated difficulties of prior learning-based tuning methods: inefficient exploration of high-dimensional mixed-type spaces and weak use of domain knowledge. Its central design is a two-part pipeline: an LLM-guided Mixture-of-Experts (MoE) that identifies performance-critical knobs and compresses their ranges, followed by a spatial decomposition procedure that recursively partitions the compressed space into hierarchical subspaces within which Bayesian Optimization (BO) is performed (Yan et al., 8 Sep 2025).

1. Problem setting and optimization target

MCTuner formulates database tuning over a mixed-type configuration space

$$\mathcal{X} \equiv R_1 \times R_2 \times \cdots \times R_n,$$

where $K=\{k_1,\dots,k_n\}$ denotes the tunable knobs and each $R_i$ is an interval for a continuous or integer knob or a finite set for a categorical knob (Yan et al., 8 Sep 2025). A configuration is represented as $\mathbf{x}=(x_1,\dots,x_n)\in\mathcal{X}$, and the optimization objective is expressed as either

$$\max_{\mathbf{x}\in\mathcal{X}_\text{feasible}} f(\mathbf{x}) \qquad \text{or} \qquad \min_{\mathbf{x}\in\mathcal{X}_\text{feasible}} f(\mathbf{x}),$$

depending on whether the tuning target is throughput, latency, or a composite metric (Yan et al., 8 Sep 2025).

The feasible region is defined not only by knob-specific bounds but also by hardware-derived limits and DBMS feasibility constraints. The details explicitly include cache sizes expressed as fractions of RAM, settings that become effective only after restart, and categorical values that are mutually exclusive (Yan et al., 8 Sep 2025). This formulation places MCTuner in the class of constrained black-box optimizers for expensive system evaluation.

MCTuner operationalizes the performance objective as a weighted relative improvement over the default configuration. When throughput and latency are jointly optimized, the scalar objective is

$$p(\mathbf{x}) \;=\; w_1\cdot\frac{\mathrm{tps}(\mathbf{x}) - \mathrm{tps}_\text{default}}{\mathrm{tps}_\text{default}} \;+\; w_2\cdot\frac{\mathrm{lat}_\text{default} - \mathrm{lat}(\mathbf{x})}{\mathrm{lat}_\text{default}}.$$

The weights are determined by user requirements and LLM inference. The reported workload-specific settings are $w_1\approx 1,w_2\approx 0$ for OLTP, $w_1\approx 0,w_2\approx 1$ for OLAP, and a mixed setting for HTAP; the experiments use $w_1=0.4,w_2=0.6$ for HTAP (Yan et al., 8 Sep 2025).

The framework is motivated by four stated challenges: the search is high-dimensional and mixed-type; evaluations are expensive because they require restart, warmup, and workload execution; vendor default domains contain vast ineffective regions; and prior learning-based tuners either ignore domain knowledge or rely on exploration that scales poorly (Yan et al., 8 Sep 2025).

2. LLM-guided Mixture-of-Experts for knob selection and range compression

The first major component of MCTuner is an LLM-guided MoE mechanism that performs adaptive knob selection and range shrinking before numerical optimization begins (Yan et al., 8 Sep 2025). Its starting point is multi-source knowledge collection. For each knob $k$, MCTuner gathers descriptions and recommended ranges from DBMS manuals, web pages, and LLM outputs, forming

$K=\{k_1,\dots,k_n\}$0

For continuous and integer knobs, the framework narrows bounds by intersecting ranges:

$K=\{k_1,\dots,k_n\}$1

Descriptions prioritize manuals over web sources and LLMs, while numeric ranges treat the three sources equally because manuals often provide overly broad bounds (Yan et al., 8 Sep 2025). The validated summary is written as

$K=\{k_1,\dots,k_n\}$2

The MoE itself contains seven domain experts specialized by functional area: Access Control, Query Optimization, Query Execution, Background Processes, CPU, Memory, and Disk (Yan et al., 8 Sep 2025). A Manager classifies each knob into relevant expert categories and assigns category weights $K=\{k_1,\dots,k_n\}$3 according to user requirements and task context, where the context includes DBMS, workload, and hardware. Experts are prompted with $K=\{k_1,\dots,k_n\}$4, independently produce an importance score $K=\{k_1,\dots,k_n\}$5, and justify the score. The final importance value is aggregated as

$K=\{k_1,\dots,k_n\}$6

MCTuner then selects the top-$K=\{k_1,\dots,k_n\}$7 knobs for subsequent optimization (Yan et al., 8 Sep 2025).

The prompting strategy uses Chain-of-Draft rather than verbose Chain-of-Thought so that only essential reasoning is retained. Hardware information is used to instantiate relative constraints into absolute values, as in converting “50% of RAM” into a specific bound (Yan et al., 8 Sep 2025). A concrete PostgreSQL example in the paper states that for OLAP on a machine with 128 GB RAM, the Manager prioritizes latency and assigns higher weights to Query Optimization and Disk experts for knobs such as random_page_cost, seq_page_cost, effective_io_concurrency, shared_buffers, and effective_cache_size; bounds such as [50%,75% of RAM] for effective_cache_size are instantiated to [64 GB, 96 GB] (Yan et al., 8 Sep 2025).

This stage is significant because it compresses the search space before expensive tuning trials begin. The paper explicitly notes that ranges for knobs such as random_page_cost and effective_io_concurrency are heavily compressed relative to overly broad manual defaults (Yan et al., 8 Sep 2025). A plausible implication is that MCTuner treats knowledge integration not as a post hoc explanation mechanism but as a structural prior on the optimization domain.

3. Spatial decomposition and MCTS-guided search

The second major component is described as the first spatial decomposition algorithm for this setting, coupled with localized BO (Yan et al., 8 Sep 2025). MCTuner recursively partitions the already compressed configuration space into hierarchical subspaces and focuses BO within promising leaves. Each iteration rebuilds a partition tree guided by Monte Carlo Tree Search, with Upper Confidence Bound scoring used to balance exploitation and exploration (Yan et al., 8 Sep 2025).

The end-to-end workflow begins with Latin Hypercube Sampling for the first $K=\{k_1,\dots,k_n\}$8 trials, after which the weights $K=\{k_1,\dots,k_n\}$9 are set once via the MoE Manager. For each iteration from $R_i$0 to $R_i$1, the tree is reconstructed from the root region $R_i$2. For a node $R_i$3 with region $R_i$4 and local dataset

$R_i$5

MCTuner computes the local statistics

$R_i$6

If the node is splittable, the region is decomposed into two children and the algorithm continues descending according to MCTS-UCB until a leaf is reached (Yan et al., 8 Sep 2025). The child-selection rule is reported as

$R_i$7

where $R_i$8 controls exploration, $R_i$9 is the parent visit count, and $\mathbf{x}=(x_1,\dots,x_n)\in\mathcal{X}$0 denote the child’s cumulative value and visit count (Yan et al., 8 Sep 2025).

Within the selected leaf subspace $\mathbf{x}=(x_1,\dots,x_n)\in\mathcal{X}$1, BO proposes the next configuration $\mathbf{x}=(x_1,\dots,x_n)\in\mathcal{X}$2; the configuration is evaluated; and the dataset is updated:

$\mathbf{x}=(x_1,\dots,x_n)\in\mathcal{X}$3

At termination, MCTuner returns

$\mathbf{x}=(x_1,\dots,x_n)\in\mathcal{X}$4

The paper states that one configuration is proposed and evaluated per iteration, and that the tree is rebuilt using the augmented dataset so that ineffective regions are progressively excluded (Yan et al., 8 Sep 2025).

This decomposition is not a static partitioning of the original space. Rather, it is rebuilt each iteration from current observations, so the hierarchy is adaptive to empirical response structure. This suggests that MCTuner uses tree search primarily as an allocation mechanism over subspaces, rather than as a direct constructive search over knob assignments.

4. Decomposition algorithm, local models, and mixed-type handling

A node in the decomposition tree represents a subspace $\mathbf{x}=(x_1,\dots,x_n)\in\mathcal{X}$5, and the split procedure uses local sample structure to partition $\mathbf{x}=(x_1,\dots,x_n)\in\mathcal{X}$6 into two children, $\mathbf{x}=(x_1,\dots,x_n)\in\mathcal{X}$7 and $\mathbf{x}=(x_1,\dots,x_n)\in\mathcal{X}$8 (Yan et al., 8 Sep 2025). Each sample is represented as

$\mathbf{x}=(x_1,\dots,x_n)\in\mathcal{X}$9

The clustering method depends on local sample count.

If $$\max_{\mathbf{x}\in\mathcal{X}_\text{feasible}} f(\mathbf{x}) \qquad \text{or} \qquad \min_{\mathbf{x}\in\mathcal{X}_\text{feasible}} f(\mathbf{x}),$$0, where $$\max_{\mathbf{x}\in\mathcal{X}_\text{feasible}} f(\mathbf{x}) \qquad \text{or} \qquad \min_{\mathbf{x}\in\mathcal{X}_\text{feasible}} f(\mathbf{x}),$$1 empirically, MCTuner uses cosine-similarity spectral clustering. The cosine similarity is

$$\max_{\mathbf{x}\in\mathcal{X}_\text{feasible}} f(\mathbf{x}) \qquad \text{or} \qquad \min_{\mathbf{x}\in\mathcal{X}_\text{feasible}} f(\mathbf{x}),$$2

and the sharpened similarity matrix is

$$\max_{\mathbf{x}\in\mathcal{X}_\text{feasible}} f(\mathbf{x}) \qquad \text{or} \qquad \min_{\mathbf{x}\in\mathcal{X}_\text{feasible}} f(\mathbf{x}),$$3

The normalized Laplacian is

$$\max_{\mathbf{x}\in\mathcal{X}_\text{feasible}} f(\mathbf{x}) \qquad \text{or} \qquad \min_{\mathbf{x}\in\mathcal{X}_\text{feasible}} f(\mathbf{x}),$$4

with $$\max_{\mathbf{x}\in\mathcal{X}_\text{feasible}} f(\mathbf{x}) \qquad \text{or} \qquad \min_{\mathbf{x}\in\mathcal{X}_\text{feasible}} f(\mathbf{x}),$$5 the degree matrix. The two smallest-eigenvalue eigenvectors provide an embedding, and K-Means with $$\max_{\mathbf{x}\in\mathcal{X}_\text{feasible}} f(\mathbf{x}) \qquad \text{or} \qquad \min_{\mathbf{x}\in\mathcal{X}_\text{feasible}} f(\mathbf{x}),$$6 is then applied in the embedded space (Yan et al., 8 Sep 2025).

If $$\max_{\mathbf{x}\in\mathcal{X}_\text{feasible}} f(\mathbf{x}) \qquad \text{or} \qquad \min_{\mathbf{x}\in\mathcal{X}_\text{feasible}} f(\mathbf{x}),$$7, MCTuner instead uses Kernel PCA plus K-Medoids. The kernel is radial basis:

$$\max_{\mathbf{x}\in\mathcal{X}_\text{feasible}} f(\mathbf{x}) \qquad \text{or} \qquad \min_{\mathbf{x}\in\mathcal{X}_\text{feasible}} f(\mathbf{x}),$$8

After centering and decomposition, the method retains the top $$\max_{\mathbf{x}\in\mathcal{X}_\text{feasible}} f(\mathbf{x}) \qquad \text{or} \qquad \min_{\mathbf{x}\in\mathcal{X}_\text{feasible}} f(\mathbf{x}),$$9 principal components with $$p(\mathbf{x}) \;=\; w_1\cdot\frac{\mathrm{tps}(\mathbf{x}) - \mathrm{tps}_\text{default}}{\mathrm{tps}_\text{default}} \;+\; w_2\cdot\frac{\mathrm{lat}_\text{default} - \mathrm{lat}(\mathbf{x})}{\mathrm{lat}_\text{default}}.$$0, and the resulting embeddings are clustered with K-Medoids (Yan et al., 8 Sep 2025).

To convert the clusters into geometric subspaces, MCTuner fits a soft-margin SVM using labels $$p(\mathbf{x}) \;=\; w_1\cdot\frac{\mathrm{tps}(\mathbf{x}) - \mathrm{tps}_\text{default}}{\mathrm{tps}_\text{default}} \;+\; w_2\cdot\frac{\mathrm{lat}_\text{default} - \mathrm{lat}(\mathbf{x})}{\mathrm{lat}_\text{default}}.$$1, solving

$$p(\mathbf{x}) \;=\; w_1\cdot\frac{\mathrm{tps}(\mathbf{x}) - \mathrm{tps}_\text{default}}{\mathrm{tps}_\text{default}} \;+\; w_2\cdot\frac{\mathrm{lat}_\text{default} - \mathrm{lat}(\mathbf{x})}{\mathrm{lat}_\text{default}}.$$2

The $$p(\mathbf{x}) \;=\; w_1\cdot\frac{\mathrm{tps}(\mathbf{x}) - \mathrm{tps}_\text{default}}{\mathrm{tps}_\text{default}} \;+\; w_2\cdot\frac{\mathrm{lat}_\text{default} - \mathrm{lat}(\mathbf{x})}{\mathrm{lat}_\text{default}}.$$3 side defines $$p(\mathbf{x}) \;=\; w_1\cdot\frac{\mathrm{tps}(\mathbf{x}) - \mathrm{tps}_\text{default}}{\mathrm{tps}_\text{default}} \;+\; w_2\cdot\frac{\mathrm{lat}_\text{default} - \mathrm{lat}(\mathbf{x})}{\mathrm{lat}_\text{default}}.$$4, which has the higher average $$p(\mathbf{x}) \;=\; w_1\cdot\frac{\mathrm{tps}(\mathbf{x}) - \mathrm{tps}_\text{default}}{\mathrm{tps}_\text{default}} \;+\; w_2\cdot\frac{\mathrm{lat}_\text{default} - \mathrm{lat}(\mathbf{x})}{\mathrm{lat}_\text{default}}.$$5, and the other side defines $$p(\mathbf{x}) \;=\; w_1\cdot\frac{\mathrm{tps}(\mathbf{x}) - \mathrm{tps}_\text{default}}{\mathrm{tps}_\text{default}} \;+\; w_2\cdot\frac{\mathrm{lat}_\text{default} - \mathrm{lat}(\mathbf{x})}{\mathrm{lat}_\text{default}}.$$6 (Yan et al., 8 Sep 2025). Along a path from the root to a leaf, the resulting region is the intersection of SVM halfspaces. The stopping rule for splitting is qualitative rather than asymptotic: a node is splittable only if the clustering yields two clusters with separable structure as judged by SVM margin, sample count, and stability across iterations (Yan et al., 8 Sep 2025).

Within a selected leaf, BO is performed with a Gaussian Process surrogate by default. Given observations $$p(\mathbf{x}) \;=\; w_1\cdot\frac{\mathrm{tps}(\mathbf{x}) - \mathrm{tps}_\text{default}}{\mathrm{tps}_\text{default}} \;+\; w_2\cdot\frac{\mathrm{lat}_\text{default} - \mathrm{lat}(\mathbf{x})}{\mathrm{lat}_\text{default}}.$$7 and responses $$p(\mathbf{x}) \;=\; w_1\cdot\frac{\mathrm{tps}(\mathbf{x}) - \mathrm{tps}_\text{default}}{\mathrm{tps}_\text{default}} \;+\; w_2\cdot\frac{\mathrm{lat}_\text{default} - \mathrm{lat}(\mathbf{x})}{\mathrm{lat}_\text{default}}.$$8, the posterior mean and variance at $$p(\mathbf{x}) \;=\; w_1\cdot\frac{\mathrm{tps}(\mathbf{x}) - \mathrm{tps}_\text{default}}{\mathrm{tps}_\text{default}} \;+\; w_2\cdot\frac{\mathrm{lat}_\text{default} - \mathrm{lat}(\mathbf{x})}{\mathrm{lat}_\text{default}}.$$9 are

$w_1\approx 1,w_2\approx 0$0

The paper lists Expected Improvement and Upper Confidence Bound as common acquisition functions:

$w_1\approx 1,w_2\approx 0$1

Alternative BO backends including SMAC, TurBO, and HEBO are reported in ablations, but the central design choice is that BO is localized to the selected subspace and coordinated globally by MCTS-UCB (Yan et al., 8 Sep 2025).

Categorical knobs are encoded numerically to allow unified surrogate modeling with continuous knobs. The paper states that categorical knobs are “encoded into continuous representations” before BO but does not mandate a specific encoding scheme (Yan et al., 8 Sep 2025). Feasibility is enforced through validated ranges from the MoE stage, hardware-aware instantiation of relative bounds, and documented DBMS operational constraints. The paper does not specify specialized kernels for categorical variables.

5. Empirical evaluation and reported performance

The reported evaluation uses PostgreSQL v14.9 and eight workloads spanning OLAP, OLTP, and HTAP (Yan et al., 8 Sep 2025). The benchmarks are TPC-H and JOB for OLAP; TPC-C, Twitter, YCSB, and SmallBank for OLTP; and CH-benCHmark and HyAdapt for HTAP. The hardware is a 24-core Intel Xeon E5-2676 v3, 128 GB RAM, 1 TB SSD, and GeForce RTX 2080 Ti with 22 GB memory (Yan et al., 8 Sep 2025). The budget is 100 iterations, with one configuration evaluated per iteration and 10 LHS cold-start trials for BO-based methods (Yan et al., 8 Sep 2025).

The baseline set includes GPTuner, ResTune, LlamaTune, SMAC, DB-BERT, CDBTune, and OpAdviser (Yan et al., 8 Sep 2025). Relative improvements are reported in latency for OLAP, throughput for OLTP, and a weighted latency-throughput objective for HTAP (Yan et al., 8 Sep 2025).

The principal quantitative outcomes are stated as follows. In latency-oriented results, MCTuner exceeds the best baseline by 6.6% on TPC-H, 5.4% on JOB, 19.2% on CH-benCHmark, and 2.4% on HyAdapt (Yan et al., 8 Sep 2025). In throughput-oriented results, the gains are 9.4% on TPC-C, 6.2% on Twitter, 12.3% on YCSB, 4.6% on SmallBank, and 13.6% on HyAdapt (Yan et al., 8 Sep 2025). Across the full set of workloads, the abstract summarizes the outcome as up to 19.2% performance gains and 1.4x faster configuration discovery per iteration than state-of-the-art methods (Yan et al., 8 Sep 2025).

The paper also reports algorithmic overhead. Excluding workload execution, average algorithm time per iteration is 5.54 s, compared with 7.47 s for GPTuner and 8.22 s for ResTune; it is slower than SMAC at 1.87 s and LlamaTune at 2.06 s, but the paper attributes substantially higher final performance to MCTuner (Yan et al., 8 Sep 2025). The space-compression analysis states that the MoE compresses many knob ranges heavily and that spatial decomposition further prunes the MoE-compressed space to a small fraction, with the reported ratio2 often much smaller than 1 (Yan et al., 8 Sep 2025).

Component-wise ablations are used to argue necessity and complementarity. On TPC-H, the spatial decomposition component contributes a 39.8% gain and the MoE contributes a 12.1% gain versus a zero-shot LLM baseline (Yan et al., 8 Sep 2025). A separate BO-backend ablation states that once the search has been restricted to productive subspaces, differences among BO choices largely vanish and overall improvements exceed 50–65% on selected workloads (Yan et al., 8 Sep 2025).

The paper also studies workload drift. Under cyclic TPC-C scale factor changes $w_1\approx 1,w_2\approx 0$2, warm-starting from top configurations of the previous phase yields immediate gains, including 10.5% for $w_1\approx 1,w_2\approx 0$3 and 20.4% for $w_1\approx 1,w_2\approx 0$4, followed by further gains under continued tuning (Yan et al., 8 Sep 2025).

6. Relation to prior tuning methods, limitations, and naming ambiguity

MCTuner is positioned against several families of system tuners. Relative to transfer-learning approaches such as OtterTune, it is described as adding MoE-driven knowledge injection to tighten ranges upfront and a decomposition stage that actively prunes ineffective subspaces during search (Yan et al., 8 Sep 2025). Relative to BO-based tuners such as iTuned, SMAC, GPTuner, ResTune, and LlamaTune, its contribution is stated as a combination of dimensionality reduction through MoE and effective-volume reduction through spatial decomposition (Yan et al., 8 Sep 2025). Relative to RL methods such as CDBTune, QTune, WATuning, UDO, HUNTER, and DB-BERT, it is described as replacing aggressive global exploration with MCTS-UCB-gated exploration and lightweight local BO (Yan et al., 8 Sep 2025). Relative to LLM-based tuners such as GPTuner, DB-BERT, E2ETune, and $w_1\approx 1,w_2\approx 0$5-Tune, the paper presents its MoE as multi-stage, interpretable, and tailored to workload, DBMS, and hardware context (Yan et al., 8 Sep 2025).

The paper does not claim formal convergence guarantees. It explicitly states that no explicit computational complexity bounds are given, and that per-iteration cost is dominated by clustering and SVM training on the current node’s samples (Yan et al., 8 Sep 2025). The framework’s stated limitations include reliance on the quality of manual, web, and LLM knowledge; imperfect treatment of semantic distances among categorical knobs under numeric encoding; added clustering overhead; sparse-data instability in early iterations; and the fact that the evaluation is limited to PostgreSQL, even though porting to other DBMSs is described as straightforward in principle (Yan et al., 8 Sep 2025). The implementation uses Python and GPT-4/gpt-4o-mini, but the paper does not claim public code release (Yan et al., 8 Sep 2025).

A common source of confusion is that “MCTuner” is not unique in the arXiv record. The term also appears as a shorthand for magnetically tunable modulation transfer spectroscopy, where it denotes Zeeman-tunable MTS for repositioning a laser-locking error signal across the $w_1\approx 1,w_2\approx 0$6 $w_1\approx 1,w_2\approx 0$7 spectrum (So et al., 2019). In a different systems context, the blockchain tuning framework BCTuner explicitly frames itself as an MCTS-based tuner and connects its design to a broader “MCTuner” concept built around LLM-guided exploration and tree-based search under high evaluation cost (Deng et al., 22 May 2026). These usages are methodologically unrelated to the PostgreSQL framework except at the level of name similarity. Within database systems, however, MCTuner specifically refers to the LLM-guided MoE plus spatial decomposition architecture introduced for database knob tuning in 2025 (Yan et al., 8 Sep 2025).