1484 - 使用检索增强生成微调 LLM 模型缓解运行时 Java 锁争用.pdf

1、Orlando,FLOctober 69IBM TechXchange 20251484Gen AI and AgentsRamiro LiscanoProfessor EmeritusOntario Tech UniversityMitigating run-time Java lock contention using a Retrieval Augmented Generation Fine-Tuned LLM ModelAcknowledgementsCoCo-investigatorinvestigatorDr.Akramul AzimDr.Akramul AzimOntario T

2、ech UniversityOntario Tech UniversityResearch TeamResearch TeamAshadullah Shawon Ashadullah Shawon MastersMastersMd Asif Khan Md Asif Khan Post DocPost DocAnkita Mukherjee Ankita Mukherjee DoctoralDoctoralIBM CollaboratorsIBM CollaboratorsVijay Sundaresan Vijay Sundaresan Cloud Cloud Performance Arc

3、hitectPerformance ArchitectYeeYee-Kang Chan Kang Chan Cloud Cloud Software ArchitectSoftware ArchitectIBM TechXchange 2025Most of these slides cover material from the following paper:Most of these slides cover material from the following paper:Shawon,A.,Liscano,R.,Azim,A.,Sundaresan,V.,&Chang,Y.K.(2

4、025,May).Shawon,A.,Liscano,R.,Azim,A.,Sundaresan,V.,&Chang,Y.K.(2025,May).Retrieval Augmented Generation FineRetrieval Augmented Generation Fine-Tuned LLM Model for Code Tuned LLM Model for Code Recommendations to Mitigate Lock Contention.InRecommendations to Mitigate Lock Contention.InCompanion of

5、the 16th Companion of the 16th ACM/SPEC International Conference on Performance EngineeringACM/SPEC International Conference on Performance Engineering(pp.95(pp.95-102).102).L Lock Contention and Performance?3Multiple threads try to acquire the same lock at the same timeMultiple threads try to acqui

6、re the same lock at the same timeIncreased waiting timeIncreased waiting timeDecreased overall performanceDecreased overall performanceThread 1Thread 1Thread 2Thread 2Lock Lock Performance IssuesPerformance IssuesTechnical DebtTechnical DebtIBM TechXchange 2025Motivation4We are focused on lock conte