The Case for Dynamic Placement of Distributed Systems Components

Abstract : Today distributed systems are made of many software components with complex interactions. One of the key challenges in such an environment is determining how to place the components so that the system performs efficiently. In this paper, we illustrate the importance of component placement with a case study, examining the performance of a common stream processing pipeline comprising Kafka, Spark, and Cassandra. We study three applications (word count, Twitter sentiment analysis, machine learning) and three placement strategies. Our results show that (i) placement has a significant impact on the application throughput (up to 52%) and (ii) the placement achieving best results differs depending on the application. We discuss why existing solutions for performance troubleshooting in distributed systems are not sufficient to help choosing an efficient placement or to detect if a chosen placement is significantly under-performing compared to others. Finally, we describe research directions to address this open problem.
Complete list of metadatas

http://hal.univ-grenoble-alpes.fr/hal-01885794
Contributor : Renaud Lachaize <>
Submitted on : Sunday, November 11, 2018 - 5:12:41 PM
Last modification on : Friday, February 15, 2019 - 12:02:49 PM
Long-term archiving on : Tuesday, February 12, 2019 - 12:31:21 PM

File

ALSAYASNEH-Compas2018.pdf
Files produced by the author(s)

Identifiers

  • HAL Id : hal-01885794, version 2

Collections

Citation

Maha Alsayasneh, Vivien Quéma, Renaud Lachaize. The Case for Dynamic Placement of Distributed Systems Components. Conférence d’informatique en Parallélisme, Architecture et Système 2018 (COMPAS 2018), Jul 2018, Toulouse, France. ⟨hal-01885794v2⟩

Share

Metrics

Record views

57

Files downloads

27