Enabling Interactive Analysis of Performance Measurements
Much of systems research consists of performance analysis -- to learn when one system outperforms another, to identify architectural choices responsible for the difference, or to identify performance anomalies in particular workloads, for example.
Performance analysis of database systems often involves running repeated experiments with varying parameters, gathering large amounts of traces or logs for analysis. Extracting insightful information from these performance measurements can be a difficult task. On one hand, the difficulty arises due to the complexity of the systems under test. In-depth expertise in each system is often essential, and no less important is a sound understanding of the underlying platform, like the OS or hardware, and how that may influence the performance of the systems. On the other hand, performance data in and of itself can present a number of challenges that are commonly encountered in large, high-dimensional data sets, such as missing or dirty data, noise, evolving set of features, unclear relationship between variables, and many others.
The Viska project aims to address these challenges by providing a new toolkit for systems researchers to generate and interpret performance measurement results, helping users derive meaningful and statistically sound conclusions. Viska leverages cutting-edge techniques from big data analytics and data visualization to aid and automate this analysis.
This site contains an initial prototype of the visualization application that is part of the Viska toolkit. Check back later for updates.