CAPES

• DOI: 10.1145/3126908.3126951
• Publication Date: 2017-11-08T21:02:30Z
• Authors: Yan Li; Kenneth Chang; Oceane Bel; Ethan L. Miller; Darrell D. E. Long
• Keywords: Lustre (file system); Benchmark (surveying); Deep Neural Networks

Parameter tuning is an important task of storage performance optimization. Current practice usually involves numerous tweak-benchmark cycles that are slow and costly. To address this issue, we developed CAPES, a model-less deep reinforcement learning-based unsupervised parameter system driven by neural network (DNN). It designed to find the optimal values tunable parameters in computer systems, from simple client-server large data center, where human can be costly often cannot achieve performance. CAPES takes periodic measurements target system's state, trains DNN which uses Q-learning suggest changes current values. minimally intrusive, deployed into production collect training actions during daily operation. Evaluation prototype on Lustre file demonstrates increase I/O throughput up 45% at saturation point.