<![CDATA[Purpose – This study aims to evaluate the performance, speedup, efficiency, and scalability of the Ray framework in a single-node virtualized environment for CPU-bound audio feature extraction tasks.Methods – An empirical benchmarking approach was employed using a dataset of 1,000 audio files with durations of 3–5 seconds. Multiple feature extraction techniques, including MFCC, spectral centroid, spectral rolloff, chroma features, and zero-crossing rate, were implemented using the Librosa library. Performance was evaluated by comparing serial and parallel execution times across different worker configurations.Findings – The results show that execution time decreased from 59.62 seconds in serial execution to 9.86 seconds when using 16 parallel workers, achieving a maximum speedup of 5.98. The system exhibits sub-linear scalability, with efficiency decreasing as the number of workers increases due to task scheduling overhead, resource contention, and virtualization constraints. An optimal performance range is observed at 8–12 workers, where significant speedup is achieved with relatively better efficiency.Research implications – This study demonstrates that the Ray framework challenges the assumption of linear scalability in CPU-bound parallel workloads by revealing how system-level constraints in virtualized single-node environments fundamentally shape speedup and efficiency trade-offs.Conclusion – This study demonstrates that the Ray framework is an effective and practical solution for accelerating embarrassingly parallel, CPU-bound workloads in single-node virtualized environments. While performance improves with increasing parallelism, careful selection of the number of workers is necessary to balance speedup and efficiency. However, the findings are limited by the use of a single-node setup and a relatively small dataset, suggesting that further evaluation in larger-scale or distributed environments is needed.]]>
Copyrights © 2026
