A 5.8 GBPS UNIFORM MAPPING DATA CENTER SWITCH Wei-Chih Lai and Ching-Te Chiu Department of Computer Science, National Tsing Hua University, Hsinchu, Taiwan ABSTRACT With the growing of cloud computing, the need of computing power no longer can be satisfied with a few powerful servers or small scale parallel computer systems. More and more servers are connected together as a data center network. Then, fault tolerance becomes an import issue when building a mas- sive data center network. Currently, many researches focus on building fat-tree data center networks. In this paper, we pro- pose a data center switch with uniform mapping connection patterns to provide higher fault tolerant capability for heavy traffic load fat-tree data center networks. A 4 × 4 banyan type switch IC is demonstrated as the commodity switch for build- ing the fault tolerant fat-tree data center networks. The 4 × 4 banyan type switch IC is fabricated in 90 nm CMOS technol- ogy, and the maximum operation rate of the IC is 5.8 Gbps with only 23 ps peak-to-peak jitter. Index Terms— Data Center Network, Scalability, Load Balance, Fault Tolerance, Fat-Tree, Load Balanced Birkhoff- von Neumann Switch. 1. INTRODUCTION Recently data center networks have become popular since the cloud computing and data center services were announced. There are various cloud services and applications nowadays, such as Amazon Elastic Compute Cloud, Google web 2.0 ap- plications, many on-line anti-virus applications, etc. With the growing of cloud computing, the need of computing power no longer can be satisfied with a few powerful servers or small scale parallel computer systems. One possible solution is to bring hundreds of thousands of servers together and connect them into a data center network. According to Cisco Data Center Infrastructure 2.5 Design Guide [1], the core-aggregation-access three-tier tree archi- tecture is adopted since tree topology is simple and easy to build. However, the bandwidth of interconnections near the core tier is huge so that the tree architecture needs numer- ous high-end switches and routers. The fat-tree architecture is recommended by M. Al-Fares, et al. [2] to solve the inter- connection capacity issue, and they also build a 27,648-node fat-tree network by using only commodity switches to support their idea. The fat-tree topology is one of the most common structure to build data center networks. However, the scalability, load balance, and fault tolerance issues in the fat-tree need to be improved to meet the growth of data center networks. With the advance of semiconductor technologies nowadays, the in- terconnections and switches in the fat-tree structure can be realized by silicon chips. The interconnections and switches of fat-tree networks become chip interconnection wires and switch IPs [3]. Then, the scalability issue can be mitigated. H. S. Chueh, et al. [4] introduce one set of special connec- tion patterns with uniform mapping property to balance traffic loading in the networks. Because traffic in the fat-tree is sep- arated evenly with the uniform mapping patterns, some of the interconnections near the root switches can be reduced. This method can provide a load balanced solution with better scal- ability for fat-tree data center networks. However, without alternative paths from the source to the destination, the fault tolerant capability in the architecture is greatly decreased. To increase fault tolerant capability of the fat-tree data center networks under heavy traffic load, we propose special connection patterns in which every connection pattern does not consume all of its bandwidth even under heavy traffic. When there are broken links or faulty switches in the net- work even in heavy traffic load situations, there are available bandwidth in every connection pattern so alternative paths are available to re-route the traffic. We also propose a new fat-tree architecture for the fat-tree data center networks for support- ing the uniform mapping connection patterns since this kinds of connection patterns is designed only for independent and identically distributed (i.i.d.) uniform traffic. We adopt the load balanced Birkhoff-von Neumann switch [5] in our archi- tecture to ensure high throughput in the fat-tree data center networks, and we also provide feedback path [6] at the same time for packet reordering. Finally, we design and implement a scalable switch IC with the proposed load balanced connec- tion patterns that can be used as the commodity switch for building the fault tolerant fat-tree data center networks. The rest of this work is organized as the follows. In Sec- tion II, we present a set of connection patterns with uniform mapping property to achieve better load balance and higher fault tolerance. Then, the load balanced fat-tree architecture is presented. In Section III, measurement results of the 4 × 4 banyan type switch with uniform mapping connection pat- terns are given. Finally, a brief conclusion is given in Section IV. 1577 978-1-4673-0046-9/12/$26.00 ©2012 IEEE ICASSP 2012