Logo des Repositoriums

Auflistung nach Schlagwort "parallelization"

1 - 2 von 2
  • Konferenzbeitrag
    Increasing the Throughput of Pipe-and-Filter Architectures by Integrating the Task Farm Parallelization Pattern
    (Software Engineering 2017, 2017) Wulf, Christian; Hasselbring, Wilhelm
    The Pipe-and-Filter style represents a well-known family of component-based architectures. By executing each filter on a dedicated processing unit, it is also possible to leverage contemporary distributed systems and multi-core systems for a high throughput. However, this simple parallelization approach is not very effective when (1) the workload is uneven distributed over all filters and when (2) the number of available processing units exceeds the number of filters. In this paper, we explain how we utilize the task farm parallelization pattern in order to increase the throughput of Pipe-and-Filter architectures. Furthermore, we describe an associated modular self- adaptive mechanism which enables the automatic resource-efficient reaction on unevenly distributed workload. Finally, we refer to an extensive experimental evaluation of our self-adaptive task farm performed by us. The results show that our task farm (1) increases the overall throughput and (2) scales well according to the current workload.
  • Konferenzbeitrag
    Predicting Scaling Efficiency of Distributed Stream Processing Systems via Task Level Performance Simulation
    (Softwaretechnik-Trends Band 43, Heft 1, 2023) Rank, Johannes; Barnert, Maximilian; Hein, Andreas; Krcmar, Helmut
    Stream processing systems (SPS) are a special class of Big Data systems that firms employ in (near) real time business scenarios. They ensure low-latency processing through a high degree of parallelization and elasticity. However, firms often do not know which scaling direction: horizontally, vertically, or mixed, is the best strategy in terms of CPU performance to scale those systems. Especially in cloud deployments with a pay-per-use model and cluster sizes that can span dozens of cores and machines, firms would profit from more accurate measurement-based approaches. In this paper, we show how to predict the CPU consumption of Apache Flink for different scaling scenarios using the Palladio Component Model. Our approach models the individual streaming tasks that make up the application and parametrizes it with fine grained CPU metrics obtained by combining BPF pro filing and querying the CPU’s performance measurement unit. Through this “task-level model approach”, we can achieve highly accurate predictions, despite using a simple model and only requiring a few mea surements for parametrization. Our experiment also shows that we achieve more accurate results than an alternative approach based on regression analysis.