PaperlayerView exact revision on arXiv

Source-linked AI summary

LOCAA: An Agentic System for Automated Lossy Compressor Tuning

Khondoker Mirazul Mumenin, Dong Dai, Sheng Di, Franck Cappello, Jinzhen Wang

arXiv:2610.10487v1cs.DC

TL;DR

An EB limits numerical reconstruction error but does not directly determine application-relevant PSNR or SSIM, and its quality mapping varies with the dataset and compressor. LOCAA iteratively selects compressor configurations, evaluates their outcomes through compression-in-the-loop search, and uses an LLM to guide later trials. PSNR-driven initialization reduced total trials from 54 to 44 and average trials per configuration from 2.70 to 2.20, an 18.5% reduction.

  • Problem

    An EB limits numerical reconstruction error but does not directly determine application-relevant PSNR or SSIM, and its quality mapping varies with the dataset and compressor.

    Source: §I · p. 1

  • Method

    LOCAA iteratively selects compressor configurations, evaluates their outcomes through compression-in-the-loop search, and uses an LLM to guide later trials.

    Source: §I · p. 1

  • Results

    PSNR-driven initialization reduced total trials from 54 to 44 and average trials per configuration from 2.70 to 2.20, an 18.5% reduction.

    Source: D. Ablation Study · p. 9

  • Takeaways & Limitations

    LOCAA supports user-defined compression, quality, and performance objectives and selects configurations that satisfy reconstruction-quality constraints while optimizing the specified objective.

    Source: §VI · p. 10

    The evaluation uses a limited set of datasets, compressors, and constraints, and generalization beyond the GPT API has not been tested.

    Source: §V · p. 9

Abstract

from arXiv · show

Large-scale scientific simulations generate substantial data volumes, making lossy compression essential for reducing storage and data movement costs. However, users configure compressors through numerical error bounds (EBs) while often evaluating results using quality metrics and end-to-end performance. Because the relationship between an EB and these outcomes varies across datasets and compressors, identifying a suitable configuration typically requires exhaustive search, which can be time-consuming and computationally demanding. We present LOCAA, an Large language model-based scientific lossy compression auto-tuning agent that performs compression-in-the-loop search using tool-integrated execution, compressor-aware guidance, and persistent memory. LOCAA supports user-defined objectives and constraints without requiring a specialized search strategy. We evaluate LOCAA across three scientific applications, five compressors, and three use cases: fixed-ratio compression, compression tuning under multiple quality constraints, and compression tuning under quality and time constraints. For fixed-ratio search across twelve fields from two applications, LOCAA requires 1.98x fewer evaluation trials than binary search and 5.03x fewer than FRaZ on average. For compression ratio maximization under joint Peak Signal-to-Noise Ratio (PSNR) and Structural Similarity Index Measure constraints across six fields from the Community Earth System Model application, LOCAA reduces the average evaluation trials from 61.5 using binary search to 17, corresponding to a 72.4% reduction. Persistent memory further reduces the average number of trials by 27.9% across multiple timesteps of the same field. These results demonstrate the potential of tool-augmented LLM agents to provide flexible and efficient compressor tuning across diverse scientific data and user-defined objectives.

I. INTRODUCTION

Scientific lossy compression reduces data volumes, but choosing an error bound that preserves application-relevant quality is difficult because outcomes vary with datasets and compressors. LOCAA uses measured compression feedback to guide tuning under user-defined objectives and constraints.

I. INTRODUCTION

  • Scientific simulations generate data volumes that can force researchers to discard outputs or reduce resolution, potentially removing physical features needed for later analysis.The passage gives large-scale NYX simulations producing terabytes of data per snapshot as an example.Source: §I · p. 1
  • An EB limits numerical reconstruction error but does not directly determine application-relevant PSNR or SSIM, so users must balance data reduction against quality preservation.The mapping from EB to quality is nonlinear and depends on both dataset and compressor, requiring repeated compression, decompression, and metric evaluation.Source: §I · p. 1
  • Growing diversity in compressors and scientific datasets makes manual tuning difficult because effective configurations vary across datasets and use cases.Source: §I · p. 1
  • LOCAA adapts its search using measured compression feedback and user-defined constraints, integrating compression-in-the-loop evaluation, task-specific guidance, tools, and persistent memory.The introduction also identifies compressor-aware strategies, including PSNR-driven initialization and staircase-aware navigation for compressors such as ZFP.Source: §I · p. 1
  • LOCAA is evaluated across three applications, five compressors, and three use cases, and reaches feasible configurations with substantially fewer evaluation trials.Source: §I · p. 2

II. RELATED WORK

Prior work applies LLM agents to scientific and systems tasks, while compression tuning methods either predict outcomes or evaluate candidate error bounds directly. LOCAA retains direct evaluations but uses an LLM agent to choose subsequent actions from measured metrics.

A. Large Language Models

  • LLM agents combine language-model reasoning with tools, memory, and iterative interaction, and have been applied to scientific experimentation, chemistry, and file-system tuning.Examples named in the passage include Curie, Chem-Crow, and STELLAR.Source: A. Large Language Models · p. 2
  • Compression tuning requires interpreting user objectives, selecting a compressor and EB, invoking evaluation tools, and revising decisions from observed compression ratio, quality, and runtime.Source: A. Large Language Models · p. 2

B. Error-Bounded Lossy Compression

  • Prediction-based and transform-based compressors reduce redundancy differently, with examples including SZ-family predictors, ZFP block transforms, and SPERR wavelets.The described compressors use distinct prediction, quantization, transform, and coding approaches.Source: B. Error-Bounded Lossy Compression · p. 2
  • Existing EB-search approaches include sampling and feature-based outcome prediction, but predictions can degrade under distribution shifts or unseen compressor behavior.Direct searches measure outcomes; binary search suits monotonic responses, while FRaZ and OptZConfig address irregular responses with optimization strategies.Source: B. Error-Bounded Lossy Compression · p. 2
  • LOCAA evaluates each candidate EB through actual compression and decompression, then uses a tool-augmented LLM agent to interpret measured metrics and choose the next action.This avoids surrogate error while supporting diverse compressors, metrics, and objectives.Source: B. Error-Bounded Lossy Compression · p. 2

C. Motivation

Compression responses depend on the input field, compressor, and EB, and their patterns may be unknown before tuning. LOCAA frames tuning as sequential search using observed metrics, user requirements, and accumulated history.

C. Motivation

  • The response pattern may be irregular or non-monotonic and can vary across fields and compressors, so choosing a search method in advance or testing several methods adds trial-and-error evaluations.The motivation is to investigate an LLM agent that adapts subsequent configurations from prior observations without requiring the response pattern or a specialized search procedure to be known in advance.Source: C. Motivation · pp. 2–3

A. Problem Formulation

  • The same EB can yield different compression ratios and reconstruction quality across datasets and compressors, so an EB found for one field or compressor cannot be assumed to transfer to another.Source: A. Problem Formulation · p. 3
  • Each actual compression evaluation reveals metrics such as CR, PSNR, and SSIM, making the EB-to-metric mapping an unknown response that may be nonlinear or non-monotonic.Source: A. Problem Formulation · p. 3
  • Users specify hard metric constraints and an objective; a configuration is feasible only when it meets every constraint, and the preferred feasible choice best satisfies the objective.Source: A. Problem Formulation · p. 3
  • LOCAA treats tuning as sequential search under a finite evaluation budget, adding each evaluated EB and its metrics to the history used to select the next EB.The stated objective is to identify a suitable configuration using as few trials as possible.Source: A. Problem Formulation · p. 3
  • When several compressors are requested, LOCAA tunes them sequentially, reuses earlier search history to guide later searches, and selects the best feasible configuration overall.The workflow is presented as an overview in Fig. 1.Source: A. Problem Formulation · p. 3, Figure 1 · p. 3

B. Overall Design

LOCAA coordinates compressor tuning through a single agent that uses measured compression feedback, structured search guidance, tools, and persistent memory.

B. Overall Design

  • A single agent keeps each tuning trajectory's evolving search state available across EB selection, execution, observation, and the next decision.Each decision accounts for prior EBs and metrics, active constraints, and the remaining evaluation budget.Source: B. Overall Design · p. 3
  • Task-specific prompts guide initialization, trial-history interpretation, compressor-aware candidate selection, constraint handling, budget-aware stopping, and final selection.The prompt structures the search policy while retaining the LLM's flexibility.Source: B. Overall Design · pp. 3–4
  • With a PSNR constraint, the agent estimates an initial EB from the target quality and data range; this is an approximate starting point, not an exact feasible bound.When PSNR is absent, the agent samples a relative EB and converts it to absolute mode using the data range.Source: B. Overall Design · p. 4
  • For ZFP's staircase-shaped response, LOCAA uses multiplicative EB changes and avoids regions with previously observed metrics to cross plateaus with fewer redundant evaluations.If the target falls between observed levels, it selects the closest feasible level and reports the gap as unattainable.Source: B. Overall Design · p. 4
  • After a task, LOCAA stores its search trajectory for reuse; the lightweight text-based memory is sufficient for studied histories, while larger histories and cross-dataset transfer remain future work.A representative five-compressor record occupied 2.1 KB, or about 760 tokens.Source: B. Overall Design · p. 4
  • MCP-exposed tools support compression execution, dataset characterization, and structured memory access, and the tool set can be extended with compressors or metrics.New compressors use the compression tool, while new metrics can be added to its outputs.Source: B. Overall Design · pp. 4–5

A. Evaluation Setup

The evaluation compares LOCAA with search baselines across scientific datasets and compressors, measuring evaluation counts and testing memory reuse and design components.

A. Evaluation Setup

  • The evaluation tests baseline performance, memory reuse, and component contributions; LLM-based configurations are run five times unless otherwise stated, with average trial counts reported.Memory initialization is compared with no initialization, and ablations assess design components.Source: A. Evaluation Setup · p. 5
  • The study covers NYX, Hurricane Isabel, and CESM fields with five compressors; SPERR and ZFP use relative bounds converted from requested absolute bounds.The compressors are SZ, SZ3, SZx, SPERR, and ZFP.Source: A. Evaluation Setup · p. 5
  • For fixed-target compression ratio, LOCAA is compared with binary search and FRaZ; for CR maximization under PSNR and SSIM constraints, it is compared with Joint BS.Joint BS searches each compressor under constraints, then selects the best feasible configuration across the pool.Source: A. Evaluation Setup · p. 5
  • Search efficiency is measured by candidate EB evaluations needed to find a feasible configuration, with metric-constrained trials including compression, decompression, and quality calculation.Fixed-ratio FRaZ determines feasibility using compressed output size.Source: A. Evaluation Setup · p. 5

B. RQ1: Benefits of the Agentic Solution

The section evaluates LOCAA across three compression-tuning objectives, from target-ratio search to joint quality-constrained and end-to-end-time optimization. LOCAA uses fewer trials in the first two tasks and selects configurations by total time in the third.

Evaluation objectives

  • The evaluation covers target-CR search, CR maximization under PSNR and SSIM constraints, and end-to-end-time minimization under quality constraints.Search efficiency is measured by the number of evaluation trials needed to identify the final solution.Source: B. RQ1: Benefits of the Agentic Solution · p. 5

Use Case 2

Use Case 1

  • 3.76 average trials versus 7.44 for binary search and 18.93 for FRaZ represents approximately 1.98× and 5.03× fewer trials, respectively.Across fields, targets, and compressors, FRaZ failed to reach the target within budget in two cases.Source: B. RQ1: Benefits of the Agentic Solution · p. 6

Use Case 3

  • At 100 Mbps, LOCAA selects SZx for CESM and HURRICANE and SPERR for NYX, balancing compressor time against transfer time rather than maximizing CR alone.The evaluation uses one field per application, with data sizes from 24.72 to 512.00 MiB.Source: B. RQ1: Benefits of the Agentic Solution · p. 7

C. RQ2: Memory Evaluation

The memory evaluation tests whether tuning knowledge from earlier timesteps helps find feasible configurations for later timesteps of the same physical field. It finds fewer average trials overall, while noting that cross-field transfer was not evaluated and may require better similarity indicators.

C. RQ2: Memory Evaluation

  • Memory reuse was tested across four timesteps each of the HURRICANE pressure and cloud-ice fields, with the same settings except for prior results at initialization.The stored tuning outcomes include evaluated compressors, error bounds, and compression-quality metrics.Source: C. RQ2: Memory Evaluation · pp. 7–8
  • 27.9% fewer trials were needed on average with memory across both fields, reducing the overall mean from 5.38 to 3.88.Memory reduced the average from 5.25 to 4.25 trials for pressure and from 5.50 to 3.50 for QICE; its benefit varied by timestep.Source: C. RQ2: Memory Evaluation · p. 8
  • Cross-field memory transfer was not evaluated, and reliable indicators for such transfer remain future work because response similarity also depends on spatial structure, compressor, and error bound.Basic statistics such as value range, standard deviation, and entropy may be insufficient to identify similar responses.Source: C. RQ2: Memory Evaluation · p. 8

D. Ablation Study

The ablation study evaluates CoT, compressor-specific knowledge, and PSNR-driven initialization as contributors to search efficiency. Compressor knowledge alone yields a larger average trial reduction than CoT alone, while initialization improves aggregate efficiency without increasing trials in any tested configuration.

D. Ablation Study

  • The CoT and compressor-knowledge ablation tested all four combinations on six CESM fields, with PSNR-driven initialization disabled to isolate these components.The task maximized compression ratio under PSNR- and SSIM-based quality targets.Source: D. Ablation Study · pp. 8–9
  • Both configurations reached PSNR 62 ± 4% on CESM FLDSC with SPERR in three trials, illustrating a case with no trial-count advantage.The authors note that approximate initialization and variation in search trajectories mean starting closer to the target does not always reduce trials.Source: Figure 4 · p. 9, D. Ablation Study · p. 9
  • 18.8% fewer average trials came from compressor knowledge alone, compared with 13.1% for CoT alone; combining them reduced trials by 15.3%, less than knowledge alone.Knowledge alone reduced ZFP trials from 8.67 to 4.00, a 53.8% reduction (2.17× speedup); the combined configuration averaged 4.97 trials versus 4.77 with knowledge alone.Source: D. Ablation Study · p. 9
  • PSNR-driven initialization estimates an initial error bound from the target PSNR, while the comparison agent starts from a random error bound.Compressor-specific knowledge was removed from both prompts to isolate initialization.Source: D. Ablation Study · p. 9
  • 18.5% fewer trials were required overall with PSNR-driven initialization, lowering the mean from 2.70 to 2.20 trials per configuration.Initialization required fewer trials in 10 of 20 configurations and matched the baseline in the other 10, without increasing trials in any configuration.Source: Table V · p. 9, D. Ablation Study · p. 9

E. Token-Usage Analysis

The token-usage analysis measures LLM interactions and token volume for tuning tasks across NYX, HURRICANE, and CESM. A run averaged 22.3 interactions, with most input-token volume served from the prompt cache.

E. Token-Usage Analysis

  • 22.3 LLM interactions were required on average for a complete tuning run across three evaluated fields, one from each application.The tasks evaluated five candidate compressors under SSIM and PSNR constraints to maximize compression ratio.Source: E. Token-Usage Analysis · p. 9
  • 88.4% of input-token volume was cached, with each run using approximately 215.9K tokens: 205.4K input and 10.5K output.Each interaction processed approximately 9.2K input tokens and generated 0.47K output tokens; approximately 181.5K input tokens were cached.Source: E. Token-Usage Analysis · p. 9

V. LIMITATIONS AND FUTURE WORK

LOCAA’s tuning requires compression-in-the-loop evaluations and LLM reasoning, adding computational overhead, cost, and latency. The conclusion presents it as a flexible agent for user-defined objectives and constraints, while calling for broader evaluation.

V. LIMITATIONS AND FUTURE WORK

  • Compression-in-the-loop tuning requires compressing and decompressing each candidate, so repeated evaluations add computational overhead, especially for large datasets or expensive compressors.LOCAA also incurs the inference cost and latency of LLM-based reasoning.Source: §V · p. 9
  • The implementation uses only the GPT API, and the evaluation covers a relatively small set of datasets, compressors, and constraints; generalization and production-scale robustness remain unestablished.Future work includes larger datasets, broader compressor collections, more complex constraints, and production-scale HPC environments.Source: §V · p. 9

VI. CONCLUSION

  • The conclusion presents LOCAA as an adaptive agent that searches compressor and error-bound configurations against user-defined compression, quality, and performance objectives.The evaluation reports that it selects configurations satisfying reconstruction-quality constraints while optimizing the specified objective.Source: §VI · pp. 9–10
Loading 2610.10487v1…