| Abstract: | The exponential growth of data science, cloud computing, distributed systems, ubiquitous computing, Internet of Things and other emerging applications and services require mega data centers to house hundreds and thousands of computing devices used to store, manipulate, analyze data and disseminate information. These massive number of computing devices are interconnected together through the data center network. The data center network is the core of the data center and it plays a pivotal role in the data center as it interconnects the whole computing devices together and basically the data center network was deploying in Three-tier architecture. The Three-tier data center network architecture is typical designed in a hierarchical tree-based structure which has three layers i.e. core, aggregation and access layers. This architecture utilizes high-end enterprise-class network devices at the core and aggregation layer. The expensive and power consumption characteristics of this high-end components are the challenges while deploying Three-tier architecture. In response to this many research efforts were done, and different types of data center network architectures were proposed. However, these proposed architectures designed to address one or two data center requirements but introduces other challenges.
Some of the proposed architectures are cost effective and less energy efficient since they use low-cost commodity off-the-shelf network devices, but they suffer incremental expansion, cabling complexity, management and operational difficulties. Some of the architectures can support large number of servers but produces poor performance network and can be complex to manage and maintain. Even if many researchers propose different data center network architectures, the Three-tier data center network architecture becomes the de-facto and most promising architecture deployed in the data center currently by data center operators. Since the cost of high-end devices used on the core and aggregation layer is one-time cost because the end-of-life of such devices is long term and also these devices replaced with low power usage chassis-based devices, still the Three-tier data center network architecture inherits low network performance specially for inter-rack traffic flows (the traffic generated by nodes in one rack or pod and flows to nodes which are installed in other rack or pod).
xi | P a g e
There are many causes that affect network performance, but the main cause of poor network performance is the oversubscription ratio. The oversubscription ratio is the ratio of downlink bandwidth to uplink bandwidth and 1:1 oversubscription ratio means all servers can communicate with their full provisioned bandwidth. In fact, this ratio can be achieved at the access layer where intra-rack traffic flow takes place i.e. the traffic stays within the rack, but moving up, to the upper layers achieving 1:1 oversubscription ratio is impossible. In this paper, the Three-tier data center network architecture is simulated using an NS-3 network simulator by changing the network structure, increasing the number of nodes and the traffic flows generated by each server in order to observe the effect of oversubscription ratio on average packet delay and throughput. This study also discusses some major network components which highly affects the performance of Three-tier data center network architecture and proposes basic solution for these issues. |
