Search, Tune, Repeat: The Search-Based Approach to Perfecting Cloud Performance

Table of Links

Abstract and 1 Introduction

1.1 Configuration Parameter Tuning Challenges and 1.2 Contributions

2 Tuning Objectives

3 Overview of Tuning Framework

4 Workload Characterization and 4.1 Query-level Characterization

4.2 Runtime-based Characterization

5 Feature Pruning and 5.1 Workload-level Pruning

5.2 Configuration-level Pruning

5.3 Summary

6 Knowledge from Experience

7 Configuration Recommendation and 7.1 Bayesian Optimization

7.2 Neural Network

7.3 Reinforcement Learning

7.4 Search-based Solutions

8 Experimental Setting

9 Related Work

10 Discussion and Conclusion, and References

7.4 Search-based Solutions

Search-based methods iteratively search for optimal configuration parameters by evaluating the system’s performance under different configurations. The core of search-based methods is to design the search strategy to handle the tradeoff between exploration(searching broadly across the configuration space to discover new and potentially promising regions) and exploitation (intensively searching in regions believed to be promising based on the available information).

BestConfig [10] recommends configuration parameters through an iterative process consisting of the divide & diverge sampling for sample collection and the recursive bound & search for configuration parameter recommendation. Specifically, it first discretizes the parameter ranges for each parameter into several intervals (divide), and then selects a set of samples by considering each interval of a parameter once (diverge). In the recursive bound & search phase, it starts by searching the point with the best performance among the selected samples. Then, it searches from its surrounding points for better performance (bound). If such point are found, it conducts sampling around the new sample. This optimization process is iteratively conducted until no point with better performance is found.

AutoTune [11] is an automatic parameter tuning system that aims to optimize application execution time on big data analytics frameworks. It executes experiments on both the production cluster and on a smaller-scale testbed to perform more test runs. It consists of three main steps. First, AutoTune executes the application using different data sizes and numbers of machines to build a parametric model. The model is used to select the data and cluster size for running the experiments within a time budget. The second step involves (i) exploration via using latin hypercube sampling (LHS) for generating different configurations to execute and (ii) exploitation via building and employing a Random Forest model for finding promising configurations to execute. In the final step, the best q configurations are used for executing the application on the production and determining the best one.

UDAO [20] considers the tuning problem as a principled multi-objective optimization (MOO) approach that takes multiple, possibly conflicting, objectives, computes a Pareto optimal set of configurations (i.e., not dominated by any other configuration in all objectives), and returns one from the set that best suits the objectives. To be specific, it incrementally transforming a MOO problem to a set of constrained optimization (CO) problems, where each CO problem can be solved individually to return a Pareto optimal point.