CSCI 8970 – Colloquium Series – Fall 2010 – Seventh Event
Modular Data Center Network Design and Implementation
Monday, October 25, 2010
Presenter: Chuanxiong Guo, Microsoft Asia
Last Monday, the University of Minnesota had the privilege of receiving Chuanxiong Guo, a well respected computer scientist from Microsoft Research Asia. Chuanxiong Guo work focuses on Data Centers and their networking complications. Currently he is focused on DCN architecture and topology design, and in particular on the Dcell, Bcube, MDCube and technologies such as DNC Management, visualization, protocol. He is currently working with a large number of graduate interns.
His talk focused on Modular Data Center Network Design and Implementation, the Bcube and MDCube, and the DNC platform (SeverSwitch). He started his talk with some background information, including the recent growth of Data Centers as a key infrastructure component for cloud computing. The USA is currently building a lot of data centers around the world. Some of which are hundreds of times larger than super-computer. The average cost is $500 million for constructing a data center. Data centers are composed of servers which are stored in rack (40 to 80 servers) which use a switch to connect all the servers.
Microsoft has currently been working on improving the way switches and servers communicate in an effort to make the communication and the transfer of data faster and more efficient. One of the ways in which they are making data centers more efficient is through the use of container-based data centers, where they can store 1000 – 2000 servers in a single container. The core benefits of shipping containers are their easy and quick deployment, high mobility and their increased cooling efficiency.
Chuanxiong Guo also discussed modular, mega-data center networking. He overviewed ways in which to connect to the server using the container: through a tree (oversubscribed, bottlenecked root), or a fat tree (1:1 subscription). While the fat tree is better, it is more expensive due to the large number of switches. As such, their goal was to create a high networked capacity, with low costs with commodity devices and low cabling complexity. By adding some switching functions to the server they can speedup for one-to-several traffic (Theorem 4), speedup for one-to-all traffic (Theorem 5), aggregate bottleneck throughput for all-to-all traffic and aggregate bottleneck through (ABT) is the total number of flows times the throughput of the bottleneck flow under the all-to-all communication pattern.
To solve the issue they created an MDCube. The MDCube leveraged the link rate hierarchy of Ethernet switches, treated each container as a “virtual node”, treated switches’ high speed uplink interfaces as “virtual interfaces”, connected switches to peer switches directly, and used virtual nodes form a data center. In conclusion, the 2-D MDCube has proved to be both cheap and effective. He also spoke of the development of a ServerSwitch and its capability as a platform, a programmable packet forwarding engine, and its flow/congestion control support. He also mentioned other related worked the MSR Asia group is researching and by doing so the audience was able to have a better understanding of some of the currently challenges faced by Microsoft through their Data Centers and some possible research studies for current graduate students at the University of Minnesota. The colloquium reminded us of the need to remember the flexibility of both hardware and software components and how one could be used solve some of the difficulties being faced with the other component.
