CS-TPRAM: A Benchmark Dataset and Framework for Microservice Response Time Prediction Under Controlled Resource Allocation and Interference

Predicting microservice response times under varying resource conditions is important for quality-of-service-aware resource management, yet no publicly available dataset combines distributed tracing with systematic resource allocation variation and controlled interference injection. This paper presents a reusable benchmarking framework and dataset for microservice response-time prediction. The framework integrates resource allocation management, interference injection, multi-source telemetry collection, and temporal synchronization. The dataset comprises 82,505 trace instances from DeathStarBench Media Microservices across 240 configurations (8 allocation schemes, 10 interference modes, 3 request rates) on a two-node Docker Swarm cluster. Using group-aware splitting, the leading models achieve $R^2$ between 0.943 and 0.945 with overlapping confidence intervals. Dynamic utilization features improve every model tested (+0.005 to +0.078 $R^2$), while memory interference and resource-constrained configurations present the greatest prediction challenges. The framework, dataset, and scripts are publicly available at Zenodo.