Abstract
In recent years, data centers have witnessed an exponential growth for hosting hundreds of thousands of servers as well as to accommodating a very large demand for resources. To fulfill the required level of demand, some approaches tackled network aspects so to host a huge number of servers while others focused on delivering rapid services to the clients by minimizing the path length between any two servers. In general, network devices are often designed to achieve 1:1 oversubscription. Alternatively, in a realistic data center environment, the average utilization of a network could vary between 5% and 25%, and thus the energy consumed by idle devices is wasted. This paper proposes a new parameterizable data center topology, called PTNet. PTNet offers a gradual scalability that interconnects small to large networks covering different ranges of sizes. This new interconnection network provides also a small path length between any two servers even in large sized data centers. PTNet does not only reduce path length and latency, it also uses a power-aware routing algorithm which saves up to 40% of energy with an acceptable computation time. In comparison to existing solutions (e.g. Flatnet, BCube, DCell and Fat-tree), PTNet shows substantial improvements in terms of capacity, robustness, cost-effectiveness and power efficiency: this improvement reaches up to 50% in some cases.
Original language | English (US) |
---|---|
Pages (from-to) | 3-18 |
Number of pages | 16 |
Journal | Journal of Parallel and Distributed Computing |
Volume | 107 |
DOIs | |
State | Published - Sep 1 2017 |
Keywords
- Average path length
- Data center network
- Energy saving
- Network topology
- Scalability
ASJC Scopus subject areas
- Software
- Theoretical Computer Science
- Hardware and Architecture
- Computer Networks and Communications
- Artificial Intelligence