Swift: Fast Performance Tuning with GAN-Generated Configurations
summary
Swift combines GAN-generated configurations with random sampling inside a Bayesian Optimization loop to skew the search toward high-performing settings, measuring generator quality with Jensen-Shannon divergence. Evaluation on Apache Flink and Spark across eight servers and an industrial workload shows other methods take 2.2–9× longer to optimize, Swift finds a 10.7× runtime speedup on one case, averages 1.28× better latency/throughput, raises CPU use by 7.5% and cuts memory from 64% to 20%, and converges much faster than vanilla BO.
AI-written summary · about
chapters (11)
- 0:00Swift performance tuning
- 1:04Configuration in big data frameworks
- 2:51Conventional ML tuning challenges
- 4:37GAN-based configuration generation
- 7:10Swift’s Jensen-Shannon divergence metric
- 9:00Swift system design overview
- 11:55Experimental setup and evaluation
- 13:43Flink benchmarks performance
- 14:46Spark benchmarks performance
- 15:53Fixed window benchmark case study
- 17:39Conclusion and future directions
related talks
34mPerformance tuning Twitter services with Graal and Machine Learning
56mPerformance tuning Twitter services with Graal and Machine Learning
01h 45mRun Frontier AI at Home
24mIntelligent system optimizations
46mPerformance-Tuning Twitter Services with Graal and Machine Learning
54mPerformance tuning Twitter services with Graal and Machine Learning
18mFast Models Need Slow Developers
23mBlazing Fast GenAI Inference With Torch.compile