acceptodds
Under review as a conference paper at ICLR 2027

MixTuner: Database-Kernel Co-Tuning via Context Transfer and Staged Search

Abstract

Automatic tuning improves database performance by replacing manual trial and error with configuration search and workload measurements. Database settings govern resource use, while operating-system kernel settings determine the mechanisms available to support it. Joint tuning can align database resource use with kernel capabilities that a single-layer search leaves fixed. Selecting these settings together requires connecting differently organized configuration knowledge and managing the cost of evaluating configuration pairs. We introduce Mixtuner, a database–kernel co-tuning framework based on context transfer and staged search. A large language model first selects database knobs and suggests numerical ranges from configuration documentation. The selected knob information and functional groups then guide Linux Kconfig menu filtering and option recommendations. After the resulting kernel is built and deployed, coarse-to-fine Bayesian optimization selects database values through measurements on that fixed kernel. The grouped context connects the configuration hierarchies, while the staged schedule shares kernel preparation across repeated database trials. Across 36 MySQL and PostgreSQL TPC-C and TPC-H scenarios, Mixtuner improves over the officially recommended configurations in 35. On MySQL TPC-H, Mixtuner also achieves lower P95 latency than the state-of-the-art database-only and kernel-only baselines.

Primary area:
infrastructure, software libraries, hardware, systems, etc.
Keywords:
Kernel Tuning, DBMS Configuration Tuning, System Optimization, Large Language Models

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_Rnx7OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

MixTuner addresses the problem that database tuning and kernel tuning have long been handled in separate layers, with single-layer search fixing the configuration of the other layer. It proposes a joint tuning framework in which an LLM selects database knobs and suggests numerical ranges from configuration documentation, guides Linux Kconfig menu filtering, and uses staged search to control the cost of evaluating configuration pairs. The work has a clear problem definition and a coherent method pipeline. The LLM is introduced to serve as a semantic pruner across heterogeneous documents rather than a direct decision-maker, which is a reasonable positioning. The manuscript is well-written and the evaluation is provided.

Critical strengths

(1) The idea of joint tuning is valuable, and the LLM is introduced to serve as a semantic pruner across heterogeneous documents. (2) The manuscript is well-orgnaized, and the evaluation seems to be solid. (3) Overall, it is a novel study.

Critical weaknesses

(1) The necessity of LLM-based pruning is expected to be discussed in detail. How much does end-to-end performance differ if pruning is replaced by rules, Bayesian optimization, or a small model?

(2) The fallback mechanism for LLM unreliability is not specified.

(3) The cost of the proposed approach needs to be quantified. The trade-off between LLM invocation/document-parsing overhead and the number of evaluations saved must be given.

(4) The latency and privacy compliance for online invocation also need discussion.

Questions for authors

see in weakness

Reviewer_nV7cOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

MixTuner addresses database–kernel configuration tuning for a fixed workload and environment, motivated by configuration-knowledge misalignment and expensive joint exploration. An LLM selects database knobs and ranges, transfers their grouped descriptions to Linux Kconfig filtering and assignment, and prepares one kernel before coarse-to-fine SMAC3 database optimization. Experiments cover MySQL and PostgreSQL, TPC-C and TPC-H, and 36 scale–terminal scenarios; method comparisons cover four tasks. The authors report improvements over reference settings in 35 scenarios and the best mean performance among evaluated methods on the four comparison tasks.

Critical strengths
  1. Novelty: The core change is conditioning kernel-menu filtering and option assignment on grouped database-knob descriptions before numerical database search, rather than independently tuning both layers. This is a novel combination: unlike GPTuner's documentation-guided coarse-to-fine database search, MixTuner also conditions kernel choices; unlike AutoOS's hierarchical kernel traversal, it supplies explicit database-knob context; unlike DB-BERT's textual tuning-hint extraction, it bridges two configuration hierarchies. In my judgment, this coordination insight could meet ICLR's novelty bar if validated; it is not a new Bayesian optimizer.
  2. A targeted context intervention: Absent, flat, and grouped context variants preserve selected knobs, ranges, and optimizer settings. Grouped context improves PostgreSQL TPC-H P95 from 3.08 to 1.80 seconds over flat context, providing evidence specifically relevant to the proposed information organization, rather than merely comparing the full system with defaults (Appendix B.5; Table 8).
  3. Explicit Reproducibility: Algorithm 1, numerical candidate/refinement rules, failure penalties, observation reuse, and prompt schemas explain how to reconstruct the workflow. The appendices also distinguish search-best observations from single measurements instead of treating all results as equivalent.
  4. An actionable performance–cost tradeoff: For operators able to rebuild kernels, the method addresses manual cross-layer tuning without additional hardware. MySQL TPC-H tuning takes 322.8 minutes and $3.90 in API expenditure, versus 632.3 minutes for GPTuner; the paper also reports faster alternatives rather than claiming universal cost superiority.
Critical weaknesses
  1. Kernel effects remain confounded — Generated kernels use Linux 6.8.12, while the reference-kernel ablation uses 6.8.0-136-generic; database settings are reoptimized, and the Reference DB kernel's identity relative to Full is unresolved (Appendices B.1, B.4). Thus Fig. 4 cannot isolate configuration-induced gains or fixed database–kernel interactions, as the appendix acknowledges. Use identical source/compiler builds and measure the four frozen database–kernel combinations with verified artifact identities; then toggle the context-induced options responsible for the improvement (Appendix B.4).
  2. Comparative search fairness is not established — MixTuner specifies 30 coarse and 50 fine history entries, with reused observations, but matched baseline evaluation limits, stopping rules, implementation provenance, and LLM settings are not reported in the main text or appendix (Appendices A.3–A.4, B.7). Table 10 gives costs, not equal-budget performance. Report performance-versus-total-time and performance-versus-new-evaluations, including failed attempts and earlier invocations, using matched backbones where applicable (Appendices B.2, B.8). This is an unresolved fairness question, not evidence that comparisons were necessarily unfair.
  3. Search-selection uncertainty exceeds the demonstrated robustness — Table 1 reports three repetitions, but independent complete-search variability remains unresolved; the 36-scenario matrix uses 32 search-best observations and four single measurements (Appendices B.1–B.2, E). Advantages over the strongest baselines are only 2.04–5.26%, making optimizer and measurement variability decision-relevant (Sec. 4.2; Table 1). Repeat complete searches with independent seeds, independently remeasure selected configurations under matched state, and report paired uncertainty for Table 1 and Table 8. Standard-deviation overlap alone does not establish insignificance (Appendix D.3).
Questions for authors

Q1. Can you provide the same-source, frozen-artifact four-cell comparison? Check if context-induced kernel changes retain the benefit.

Q2. Can you disclose baseline implementations/backbones and produce a matched-budget comparison? Check if the advantage persists after counting all attempts.

Q3. Are Table 1 repetitions independent searches or configuration measurements, and can paired independent-search/remeasurement results confirm Table 8's grouping effect?

Reviewer_Z1btOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper presents MixTuner, a framework for jointly configuring a database and its Linux kernel. An LLM selects database knobs and proposes numerical ranges from configuration knowledge. The selected knobs' descriptions and relationships, organized by database function, condition kernel-menu pruning and option assignment. The resulting kernel is built and deployed once; coarse-to-fine SMAC3 Bayesian optimization then tunes database values on that fixed kernel. The main contribution is the transfer of database-specific context across configuration hierarchies, combined with a staged execution schedule that amortizes kernel preparation. Evaluation covers MySQL and PostgreSQL on BenchBase TPC-C and TPC-H across 36 scale/concurrency scenarios. The paper reports improvements over untuned reference configurations in 35 scenarios and the best mean performance among the compared methods on four tasks at SF=1 and T=6. Context, configuration-layer, and backbone ablations provide additional evidence, with tuning-time and API-cost comparisons for one scenario.

Critical strengths
  1. A useful interface for a meaningful systems problem. Database knobs and kernel build options describe related resource-management behavior through different configuration hierarchies. Transferring selected knob descriptions and relationships supplies more actionable context than a shared performance objective alone. Combining this interface with a single kernel deployment makes the approach operationally plausible despite expensive kernel transitions.

  2. Relevant comparisons and encouraging context ablations. The paper includes composed kernel-plus-database tuners, rather than relying solely on comparisons against defaults or single-layer methods. Table 8 also compares absent, flat, and grouped database context while retaining selected knobs, ranges, and optimizer settings. The reported reductions from 6.71 to 4.87 seconds on MySQL TPC-H and from 4.21 to 1.80 seconds on PostgreSQL TPC-H are encouraging evidence for the proposed interface, although their repeatability remains unresolved.

Critical weaknesses
  1. The evidence does not establish a reliable advantage over strong tuners. In Table 1, improvements over the strongest baselines are only 2.04–2.18% in throughput and 2.60–5.26% in P95 latency. These gaps are small relative to the reported run-to-run variation; for example, MySQL TPC-H is 4.87 ± 0.49 seconds versus GPTuner's 5.00 ± 0.60 seconds, with only three repetitions. The description does not clearly distinguish independent complete tuning runs from remeasurements of selected configurations. Moreover, 32 of the 36 broader-scenario results are search-best observations and four are single measurements, so the 35/36 improvement count is vulnerable to selection bias and unquantified measurement noise. The headline 33.58× improvement over defaults is mostly reproduced by database-only tuning, which already reaches 5.00 seconds from 163.55 seconds. Consequently, the central empirical case requires independent post-selection measurements and uncertainty estimates for the differences against strong baselines.

  2. The experiments do not adequately isolate the central cross-layer contribution. The numerical search explicitly builds on existing coarse-to-fine Bayesian optimization, making the context interface particularly important to validate. However, the configuration-layer comparison uses different Linux builds (6.8.12 versus 6.8.0-136-generic) and reoptimizes database settings; Appendix B.4 additionally leaves unresolved whether the Reference DB variant uses the same generated kernel as Full. These workflow comparisons cannot isolate kernel-configuration effects or demonstrate a fixed database–kernel interaction. The context ablation is more directly relevant, but provides no uncertainty estimates across independent kernel generations and subsequent searches, nor concrete effective configuration differences explaining its gains. The existing results therefore leave uncertain whether functional context consistently produces better cross-layer decisions. This affects the strength of the paper's main contribution, rather than merely its deployment breadth.

Questions for authors
  1. Clarification and supporting evidence: Do the three Table 1 repetitions rerun the entire LLM/kernel-generation/database-search pipeline, or remeasure previously selected configurations? Can you provide the individual observations and uncertainty estimates for performance differences against the strongest baselines, and clarify whether the 36-scenario winners were independently remeasured after selection?

  2. Clarification and attribution evidence: Can you establish the kernel identities in the configuration-layer variants and provide effective kernel/database configuration differences for the absent, flat, and grouped context variants? Do the gains persist over independent runs with a common kernel source version? One controlled example linking a context-induced option change to workload behavior would materially strengthen the main claim.

  3. Comparison protocol: What evaluation counts, stopping rules, and preprocessing costs were used for each baseline, and what does Table 10 include for kernel compilation, boot/deployment, failed or superseded searches, and offline preparation? Does the performance ranking persist under a common total tuning-time budget? This matters because MixTuner takes 322.8 minutes versus SysInsight's 19.8 minutes on MySQL TPC-H for a reported 4.13% mean P95 improvement.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.