Authors:
Hawzhin Hozhabr Pour
1
;
Gabriela Ciortuz
1
;
André Lüers
2
and
Sebastian Fudickar
1
Affiliations:
1
Institute of Medical Informatics, University of Lübeck, Ratzeburger Allee 160, 23562 Lübeck, Germany
;
2
Department of Informatics, University of Oldenburg, Carl von Ossietzky Universität Oldenburg Ammerländer Heerstraße 114-118, 26129 Oldenburg, Germany
Keyword(s):
Network Architecture, Streaming Media, Performance Evaluation, Throughput, Latency, Distributed Computing, Human Activity Recognition, Pattern Recognition, Wearable Sensors.
Abstract:
Online activity recognition based on wearable sensors is commonly used in sports and medicine applications. The question of whether cloud or edge computing approaches are more suitable is not easy to answer and depends on several factors. To address this issue, the influence the resource availability, batch sizes and number of considered users on the throughput and latency of central data stream processing architectures has yet to be answered. This article conducts a performance analysis, identifying relevant factors for a corresponding cloud-based online data stream processing platform for online human activity recognition, using the Apache Spark data processing framework and the Apache Kafka distributed messaging system. The platform focuses on quantitative performance criteria to evaluate its effectiveness in terms of latency (turnaround time) and throughput (number of users). Both metrics, throughput and latency (dependent variables), depend on the batch interval, number of users
, and hardware availability (independent variables). In addition to identifying clear advantages of larger batch intervals, we also found significant benefits in applying vertical scaling. The results indicate a monthly cost of 1e per user for compute resources in online activity recognition, a price that could potentially be reduced by combining edge and cloud computing.
(More)