Network topology optimization for data aggregation with splitting

• Published in: IEEE International Symposium on Signal Processing and Information Technology pp. 000398 - 000403
• Main Authors: Das, Soham, Sahni, Sartaj
• Format: Conference Proceeding
• Language: English
• Published: IEEE 01.12.2014
• Subjects: Approximation algorithms; Big Data applications; Clustering algorithms; Complexity theory; Data Center Networks; Map-Reduce tasks; Network topology; Optical fiber communication; Optical switches; Software Defined networking; Topology
• DOI: 10.1109/ISSPIT.2014.7300622
• ISSN: 2162-7843

In this paper, we develop algorithms for the data aggregation problem which arises in the context of big-data applications that employ the MapReduce operation. For the case when source racks can send their data to the aggregator using multiple paths, we show that an aggregation tree topology that minimizes aggregation time can be constructed in polynomial time. We consider also the problem of constructing aggregation trees that minimize total network traffic subject to the primary constraint that aggregation time is minimized. Heuristics for this problem are presented. Experiments show that allowing multiple paths reduces aggregation time by up to 99% relative to the aggregation trees constructed using the LPT rule [3]. This reduction in aggregation time, however, comes with up to 35% increase in total network traffic when racks have more than 2 optical links.