Data Stream Operations as First-Class Entities in Component-Based Performance Models

Abstract

Data streaming applications are an important class of data-intensive systems. Performance is an essential quality of such systems. It is, for example, expressed by the delay of analysis results or the utilization of system resources. Architecture-level decisions such as the configuration of sources, sinks and operations, their deployment or the choice of technology impact the performance. Current component-based performance prediction approaches cannot accurately predict the performance of those systems, because they do not support the metrics that are specific to data streaming applications and only approximate the behavior of data stream operations instead of expressing it explicitly. In particular, operations that group multiple data events and thus introduce timing dependencies between different calls to the system are not represented sufficiently. In this paper, we present an approach for modeling networks of data stream operations including their parameters with the goal of predicting the performance of the resulting composed data streaming application. The approach is based on a component-based performance model with queueing semantics for processing resources. Our evaluation shows that our model can more accurately express the behavior of the system, resulting in a more expressive performance model compared to a well-encapsulated component-based model without data stream operations.

Partially funded by the German Research Foundation (DFG) as part of the Research Training Group GRK 2153: “Energy Status Data – Informatics Methods for its Collection, Analysis and Exploitation”.