Technical Presentations

Опубликовано 6 сентября 2016, 18:51

Using Machine Learning to Characterize, Predict and Optimize System Behavior - Presented by Archana Ganapathi Abstract: Systems have become very complex and it is often difficult to maintain an accurate model of various system components and their interactions. We use Statistical Machine Learning algorithms to automatically extract relationships between system workload and behavioral metrics. Using Kernel Canonical Correlation Analysis (KCCA) to extract correlations between system workload and observed system behavior, we determine clusters of similar requests that, with high confidence, cause a particular system behavior (e.g. 80 Predicting Web 2.0 Workload Behavior - Presented by Kristal Sauer Abstract: The aim of this project is to use statistical machine learning to predict a system's performance and resource utilization under changes to the workload or underlying hardware. This could be useful in many scenarios. For instance, in the web service domain, when a company wants to promote a feature of their application and thereby shift the distribution of requests towards a certain type, they would like to have a sense for the resulting performance. Also, when it comes time to upgrade their servers, they would like to predict the system's behavior on the new hardware. This work is inspired by a study done by Ganapathi et al. (ICDE '09) in which this approach was utilized for predicting query runtime and resource consumption; the contribution of this work is the application of the predictive framework to the web service domain via analysis of a new Web 2.0 social networking benchmark application called Cloudstone. In previous experiments, we validated that our representation of the input allowed us to recreate the CPU utilization of the system; in this phase of the project, we focus on predicting the performance and resource utilization of a given workload. In this talk, I will describe our framework and present experimental results.