Carver: Finding Important Parameters for Storage System Tuning
summary
Carver measures each storage parameter’s effect as the variance-based metric (reduction in response variance), samples huge spaces with Latin Hypercube Sampling, and runs a greedy conditional-selection algorithm to pick parameters that cut variance most. Evaluation uses >500,000 runs across 7 file systems and 4 workloads, finds that importance shifts with workload, and recovers a near-optimal small parameter set while using as little as 0.4% of the full dataset.
AI-written summary · about
chapters (18)
- 0:00Intro
- 0:18Outline
- 2:48Carver Overview
- 3:49Key Contributions
- 4:45Concepts
- 5:38Challenges
- 6:49Dimensionality Reduction
- 8:12Carver: Measuring Parameter Importance
- 9:07Carver: Sampling
- 9:52Carver: Parameter Selection (2)
- 12:24Evaluation: Testbed
- 13:08Evaluation: Settings
- 13:57Evaluation: Parameter Importance
- 14:56Parameter Importance Summary
- 18:32Evaluation: Carver (Cont.)
- 19:40Related Work
- 20:39Conclusions
- 21:12Future Work
related talks
46mLinux Kernel Autotuning
36mStorage Wars
18mLet the machines optimize the machines
27mSolving inefficiency challenges at Datacenter scale
36mAutoscaling Isn't Set-and-Forget: Tuning KEDA & Karpenter for Real Workloads.
10mCluster-scale Problem Solving
12mEcosystem Tools to Select the Best Remote Storage Solution
39mBig Data Systems Performance: The Little Shop of Horrors