Flow Diffusion Algorithms Based on Local and Semi-Local Information for Folded Clos Networks Satoru Ohta Department of Electrical and Computer Engineering, Faculty of Engineering Toyama Prefectural University Kurokawa 5180, Imizu-shi, Toyama, 939-0398 Japan e-mail: ohta@pu-toyama.ac.jp ABSTRACT Folded Clos networks (FCNs) are often employed as a topology for data center networks. When an FCN is applied to a data center network, traffic congestion on links must be avoided to enable high throughput data transmission. To avoid congestion, flows need to be uniformly diffused between the links. Therefore, this paper investigates two algorithms that nearly uniformly spread flows in FCNs. The first algorithm is executable with information that is locally obtainable at the source of a flow. This method is further divided into two versions. One version reroutes some existing flows, while the other does not perform rerouting. The second algorithm runs with semi-local information, which can be obtained via communication between the source and the destination switches. The processing overhead for these algorithms is small because global information is unnecessary. The paper evaluates the presented algorithms and conventional random routing. The result clearly shows the advantage of the presented algorithms. KEYWORDS network; algorithm; routing; data center; packet. 1 INTRODUCTION Nowadays, most information services are provided via servers located in data centers. Therefore, the quality of the service greatly depends on the performance of the data centers. This necessitates the establishment of technologies for high performance data center networks. Past studies have examined various topologies, such as a Clos network [1], fat-tree [2], DCell [3], and BCube [4], for data center networks. Of these topologies, the Clos network is advantageous because it provides high throughput for arbitrary traffic patterns. Therefore, various data center networks based on the Clos network topology have been implemented and are currently operating [1, 5–7]. The Clos network is a switching network, first investigated by Charles Clos [8], which provides large-scale switching function by networking multiple small switches in three stages. The first stage includes input switches, and the third stage includes output switches. Middle switches in the second stage connect the input and output switches. In data center network applications, the network is often represented in the form of a folded Clos network (FCN). An FCN is constructed by folding a Clos network at its center. When an FCN is applied to a data center network, it is important to appropriately route packets at the input switches. Inadequate routing may cause load unbalances between the links. Such an unbalance may lead to further congestion and therefore degrade the system performance. To evenly diffuse the link load, conventional implementations [6, 7, 9] randomly forward packets from input switches to middle switches. This random route selection is often performed using a hash function [7, 9], which is fed flow identifiers as input. On average, this approach achieves uniform load diffusion. However, with this method, the traffic on some links may grow excessively large with a substantial probability. As [10] pointed out, this problem may become critical for big data applications, which require high bandwidth transmission. Therefore, it is important to develop a routing algorithm that more uniformly diffuses the load. In addition, switches need to route flows without global Proceedings of The Fourth International Conference on Electronics and Software Science (ICESS2018), Japan, 2018 ISBN: 978-1-941968-52-9 ©2018 SDIWC 46