Use the fat tree topology to effectively improve the performance of the Ethernet Switch

According to researchers at the University of California, San Diego, purchasing a faster switch may be the only way to improve the network performance of the entire data center. These researchers proposed this week a network architecture that enables commercialized Ethernet switches to provide better performance at a lower cost than 10-Gigabit Ethernet switches.

Amin Vahdat, a professor of computer science at the University of California, San Diego, presented the findings at the SIGCOMM 2008 conference in Seattle, this article describes how to use the Cluster Computing Principle in the network architecture to improve the scalability and performance at a low cost.

Vahdat, one of the three authors of the paper "network architecture of a scalable and commercialized data center", said that the data center was not built based on high-end components, it is built based on the network that we still rely on high-end advanced technologies.

Vahdat says there is no need to invest in special devices, such as 10-Gigabit Ethernet switches and routers, or use a standard three-tier architecture, enterprises can use commercial Ethernet switches in a "fat tree" architecture to achieve the same performance at a very low cost. Interestingly, he explained that a network with 20 thousand nodes using an expensive switch costs $28 million. Using commercialized devices, the same network only costs nearly $4 million.

Vahdat says there are many optimistic figures. However, the advantages of commercialized products will be disruptive to computing and technology. Throughout history, there are many examples of using telephone switches and commercialized PC clusters. These examples show how commercialized components conquer an industry.

These results were proposed when a high-end data center supporting thousands of nodes emerged and broke the old network model. This will make IT managers look for a choice, because the 40 GB Ethernet standard will be available in a few months, and IT will take several years for the GB Ethernet standard to be available.

Said Phil Hochmuth, a senior analyst at Yankee Group, Ethernet technology is slowly entering the last few bastion hosts of enterprise proprietary/EDGE network technology. These bastion hosts are mainly used in data centers. Consider the fiber channel for storage, Infiniband or Myrinet Technology for applications for high-performance computing and cluster applications. The low cost and flexibility of Ethernet are driving this trend. The study at the University of California, Santiago, is a good example of how far this idea can go.

Vahdat pointed out that because of the small difference, the price reasons may not be very attractive at present. However, when 40 GB Ethernet becomes a reality, the price of each port will be very expensive and the port density will be very small.

The research team began researching high-cost, low-performance, and high-end networks about a year ago. Conversations with many companies reveal complaints about high-cost and complex data centers that lack bandwidth. Vahdat says the problem has become more serious over the past three years because more enterprises are building data centers that cannot provide promised performance and require a lot of investment. Although the cost of using the current technology cannot be considered as the main driving factor for adopting the proposed architecture, the price will become a problem in the near future.

Vahdat explains that it is not suitable for considering the price of each port yet. However, once you push 10-ge Ethernet to the edge, the existing design will not work. If you want to push 10-ge Ethernet to a server in a large cluster, you have no choice.

Vahdat and his colleagues used the fat tree topology to design an Ethernet switch interconnection method so that all the vswitch elements are identical, thus, it is possible to use cheap commercial components for all switches in the communication architecture. He explained that to use similar devices, the fat tree topology network depends on protocols that can take advantage of all the advantages of different paths in the network.

We don't think you need to modify the chip, said Vahdat. This work can be completed through the software, only a small modification is required.

The researchers suggested that the use of second-level route tables and traffic classification technologies and other technologies would allow these switches to route communication traffic in the entire fat tree chart without generating bandwidth bottlenecks. However, Vahdat says his team still has to solve two problems. The first problem is how many modifications (software modifications or other modifications) the network administrator needs to make the switch take advantage of the fat tree topology to route the communication. The second unsolved problem is the wiring problem. Currently, the network connected to this network using a high-end 10-Gigabit Ethernet switch has fewer cables than the fat tree network. The fat tree network requires 100 separate cables for 100 low-speed connections.

Vahdat explains that we are determining the minimum number of changes to the existing vswitch to implement the routing function and solve the wiring problem raised by the fat tree topology.

• • The S2200MF Gigabit Ethernet switch launched by beacon Network