Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

EEvA: Fast Expert-Based Algorithms for Buffer Page Replacement

Alexander Demin, Yuriy Dorn, Aleksandr Katrutsa, Daniil Kazantsev, Ilgam Latypov, Yulia Maximlyuk, Denis Ponomaryov

TL;DR

The paper tackles the challenge of designing buffer eviction policies that are both adaptive to diverse data-access patterns and computationally lightweight for DBMS workloads. It proposes EEvA, an expert-based eviction framework that combines static page/table-level experts with online learning (Online Mirror Descent) and a Markov Decision Process perspective to model eviction decisions. The authors derive several concrete algorithms (EEvA, EEvA-Greedy, EEvA-T, EEvA-Seq), provide regret bounds, and validate performance through synthetic experiments and an open-source DB kernel on TPC-C, showing improvements in hit rate and latency over strong baselines except in the theoretical worst-case belady scenario. The work demonstrates practical impact for large buffer pools and heavy concurrency, and points to future work on deeper theoretical analysis and more integrated, production-ready implementations of page-flush policies.

Abstract

Optimal page replacement is an important problem in efficient buffer management. The range of replacement strategies known in the literature varies from simple but efficient FIFO-based algorithms to more accurate but potentially costly methods tailored to specific data access patterns. The principal issue in adopting a pattern-specific replacement logic in a DB buffer manager is to guarantee non-degradation in general high-load regimes. In this paper, we propose a new family of page replacement algorithms for DB buffer manager which demonstrate a superior performance wrt competitors on custom data access patterns and imply a low computational overhead on TPC-C. We provide theoretical foundations and an extensive experimental study on the proposed algorithms which covers synthetic benchmarks and an implementation in an open-source DB kernel evaluated on TPC-C.

EEvA: Fast Expert-Based Algorithms for Buffer Page Replacement

TL;DR

The paper tackles the challenge of designing buffer eviction policies that are both adaptive to diverse data-access patterns and computationally lightweight for DBMS workloads. It proposes EEvA, an expert-based eviction framework that combines static page/table-level experts with online learning (Online Mirror Descent) and a Markov Decision Process perspective to model eviction decisions. The authors derive several concrete algorithms (EEvA, EEvA-Greedy, EEvA-T, EEvA-Seq), provide regret bounds, and validate performance through synthetic experiments and an open-source DB kernel on TPC-C, showing improvements in hit rate and latency over strong baselines except in the theoretical worst-case belady scenario. The work demonstrates practical impact for large buffer pools and heavy concurrency, and points to future work on deeper theoretical analysis and more integrated, production-ready implementations of page-flush policies.

Abstract

Optimal page replacement is an important problem in efficient buffer management. The range of replacement strategies known in the literature varies from simple but efficient FIFO-based algorithms to more accurate but potentially costly methods tailored to specific data access patterns. The principal issue in adopting a pattern-specific replacement logic in a DB buffer manager is to guarantee non-degradation in general high-load regimes. In this paper, we propose a new family of page replacement algorithms for DB buffer manager which demonstrate a superior performance wrt competitors on custom data access patterns and imply a low computational overhead on TPC-C. We provide theoretical foundations and an extensive experimental study on the proposed algorithms which covers synthetic benchmarks and an implementation in an open-source DB kernel evaluated on TPC-C.
Paper Structure (36 sections, 2 theorems, 21 equations, 11 figures, 3 tables, 4 algorithms)

This paper contains 36 sections, 2 theorems, 21 equations, 11 figures, 3 tables, 4 algorithms.

Table of Contents

  1. Introduction
  2. Related works
  3. FIFO-based strategies
  4. Recency-based strategies
  5. Frequency-based strategies
  6. Randomized strategies
  7. Machine Learning-based strategies
  8. Online Convex Optimization in eviction strategies
  9. Buffer management problem as online learning problem
  10. General expert-based framework
  11. Experts for page replacement
  12. Buffer Eviction Strategy Based on Expert Model
  13. Markov Decision Process model for page replacement
  14. From DB buffer model to expert-based eviction algorithms
  15. Experts and losses
  16. ...and 21 more sections

Key Result

Theorem 1

Let $\mu = \sqrt{\frac{8}{T} \log T}$, $0<\beta<\alpha\leq 1$, and EEvA algorithm outputs a sequence $\{x_t\}_{t=1}^T$. Then the following inequality holds:

Figures (11)

  • Figure 1: Expected average page weights corresponding to the scenario with only get-type queries, in which every page has the same probability of being requested. Weight values are spread uniformly along the interval of page indices.
  • Figure 2: EEvA-T and EEvA-Greedy algorithms outperform competitors in the get-only scenario, where all pages are sampled for get-type queries with similar probabilities. These plots show the robustness of the proposed algorithms in the scenario with no scan queries.
  • Figure 3: Expected average page weights $w_i$ show that queries to the first pages assigned to scanned tables are less costly than others. Since we take into account not only the frequency of the requested pages but also the corresponding cost ($c_{scan}$ or $c_{get}$), the expected average weights accumulate this information, as well.
  • Figure 4: Comparison in the scenario with rare scan-type queries. The lower the miss rate and averaged time costs, the better the performance is. EEvA-based algorithms show a superior performance over the competitors wrt both metrics.
  • Figure 5: Expected average page weights $w_i$ show that queries to the first pages assigned to scanned tables are comparable in terms of the loading costs compared to other pages. Increasing the value of $p_{scan}$ leads to increased costs of scanned pages, which affects the distribution of the shown expected average weights.
  • ...and 6 more figures

Theorems & Definitions (4)

  • Theorem 1
  • proof
  • Theorem 2
  • proof