| CHOONHAKLAI PAPON | M, 2回目発表 | ソフトウェア設計学 | 飯田 元, | 藤川 和利, | 市川 昊平, | 柏 祐太郎 |
|
title: A Kubernetes Operator Design to Enhance MPS-Based GPU Sharing
abstract: Efficient GPU sharing is essential for Kubernetes clusters running deep-learning inference workloads, as assigning an entire GPU to each pod wastes expensive hardware. This work proposes a three-step design that progressively improves sharing efficiency. Approach 1 combines NVIDIA time-slicing with KubeRay: a single job can occupy the full GPU when the cluster is idle, but additional jobs share GPU time slots as they arrive. This can cause high context-switching overhead when many jobs are submitted. Under this scheme, 75 BERT inference jobs complete in 1,302 seconds, with the GPU averaging 72% utilization and only 22.6% memory usage—indicating significant underutilization. Approach 2 replaces time-slicing with NVIDIA Multi-Process Service (MPS). A lightweight Kubernetes operator works with Nebuly’s NOS MPS server to inject fixed CUDA thread limits based on node priority, reducing total runtime to 1,006 seconds, increasing GPU utilization to 90%, and raising memory usage to 30%. Approach 3 enhances the operator with DCGM Exporter telemetry, a weighted-score scheduler, and a per-node finite-state machine that throttles, rescales, or restarts pods as load conditions change. This metric-driven method supports 36 concurrent BERT inference jobs, cuts completion time to 954 seconds, and raises GPU utilization and memory usage to 99% and 33.7%, respectively. Overall, the third approach achieves a 10.6% speed-up over NVIDIA time-slicing and is 3.1% faster than a standalone NOS configuration—approaching full GPU saturation and offering a practical path to high-throughput, policy-aware accelerator sharing in Kubernetes. language of the presentation: English | ||||||
| 稲葉 真幸 | M, 2回目発表 | ソフトウェア工学 | 松本 健一, | 飯田 元, | Raula Gaikovina Kula, | 嶋利 一真 |
|
title: Detecting Textually Dissimilar Duplicate Bug Reports using Transformers
abstract: In large-scale bug management systems, hundreds of bugs are reported daily, and it is not uncommon for multiple individuals to report the same bug, resulting in duplicate bug reports. Manually classifying duplicates places a burden on developers, prompting research into automated detection methods. However, since bug reports are written in natural language, differences in subjective interpretation, perceptions of the issue, and circumstances surrounding the bug often lead to variations in the descriptions used in reports. In this study, such cases are referred to as textually dissimilar duplicate bug reports. Detecting textually dissimilar duplicate bug reports is challenging with traditional natural language processing (NLP) or machine learning methods, as it requires accurately understanding the semantics of each report's description. This study aims to identify effective methods and models for detecting textually dissimilar duplicate bug reports. An evaluation experiment was conducted using Transformer-based large language models and Sentence Transformers, which specialize in comparing textual similarity. The results showed that Sentence Transformer-based models achieved high precision in terms of both F-score and AUC for detecting both textually similar and textually dissimilar duplicate bug reports, demonstrating their utility in detecting duplicate bug reports. language of the presentation: Japanese 発表題目: テキスト表現が異なる重複バグ報告の検出におけるTransformerモデルの利用 発表概要: 大規模なバグ管理システムでは毎日何百ものバグが報告されるため，複数の人が同じバグについて報告を行ってしまうことがあり，重複したバグ報告が発生する．重複を手動で分類することは開発者にとって負担となるため，自動的に検出する手法が研究されてきた．しかし，バグ報告は自然言語で構成されるため，報告者の主観や認識内容の相違，発生状況によって，報告の説明文に異なる表現が用いられることがある．本研究ではそれらをテキスト表現が異なる重複バグ報告と呼ぶ．テキスト表現が異なる重複バグ報告は，それぞれのバグ報告の説明文が表す意味を正しく理解しないと重複と判断することができないため，従来の自然言語処理や機械学習の手法では検出が難しい場合がある．本研究では，テキスト表現が異なる重複バグ報告を検出するための効果的な方法とモデルを特定することを目的とし，Transformerベースの大規模言語モデルと，文章の類似性比較に特化したSentence Transformersを用いて検出性能の比較を行う評価実験を行った．実験の結果，テキスト表現が類似している重複バグ報告，類似していない重複バグ報告のいずれにおいてもF値とAUCにおいてSentence Transformerベースのモデルが高い精度を示し，重複バグ報告の検出において有用であることが明らかとなった． | ||||||
| 田中 英武 | M, 2回目発表 | ソフトウェア工学 | 松本 健一, | 飯田 元, | Raula Gaikovina Kula, | 嶋利 一真 |
|
title: Applying Property-based Testing Generated by LLMs to Code Generation
abstract: In the field of code generation using large language models (LLMs), approaches that iteratively test and refine generated code to improve its correctness and quality have recently attracted attention. The primary testing method used in these approaches is Example-based Testing, which checks the correctness of code using specific input examples. However, this method lacks comprehensiveness and often fails to detect hidden bugs. To address this limitation, our study applies Property-based Testing—a technique that verifies general properties which the code should satisfy, rather than relying solely on specific examples—to code generated by LLMs. This approach enables more efficient bug detection across a broader input space. As an evaluation, we conduct tests on code generated by LLMs using both Property-based Testing and Example-based Testing, with both types of tests also generated by LLMs. By comparing the test results, we aim to clarify the effectiveness and usefulness of Property-based Testing in LLM-based code generation. language of the presentation: Japanese 発表題目: LLMを用いたProperty-based Testingのコード生成への適用 発表概要: 大規模言語モデル（LLM）を用いたコード生成において，近年精度向上を目的として生成したコードのテストと修正を反復的に行うアプローチが注目されつつある．このアプローチでは，主なテスト手法として具体的な入力例をもとにコードの正否を判断するExample-based Testingが用いられるが，網羅性に欠け，隠れたバグを検出しにくいという課題がある．本研究では，この課題への対応を目的として，LLMが生成したコードに対してProperty-based Testingと呼ばれるテスト手法を適用する．Property-based Testingは，入力例に依存するのではなく，コードが満たすべき一般的な性質をテストする手法であり，広範な入力空間に対して効率的なバグ検出が可能となる．評価として，LLMが生成したコードを対象に、同じくLLMが生成したProperty-based TestingとExample-based Testingを用いてテストを実施する．テスト結果を比較することで，LLMによるコード生成におけるProperty-based Testingの有用性を明らかにすることを目指す． | ||||||