LAN switches are "big" (1)

What is the largest and fastest LAN switch? This is actually a question with many complex answers, and the answer depends on how you define "big" and "fast ".

3Com, Extreme, Force10, Foundry, HP ProCurve, Cisco, and other Ethernet switch vendors fight for the highest performance, maximum density, and minimum latency. However, the products and technologies provided by these manufacturers have a history of three years on average. At the same time, a group of hungry emerging vendors, such as Woven Systems and Raptor Networks, are initiating new challenges in how to build the "largest" Ethernet switch. They are different from a single large chassis, but more like cluster-based supercomputing or InfiniBand Network topology.

High performance is reflected in many aspects

The Ethernet speed is limited by the current 802.3ae standard speed of 10 Gbps). Therefore, it can be inferred that the speed of no independent port exceeds 10 Gbps. Other methods to measure the performance of a vswitch are based on the bandwidth of the vswitch structure and the port density supported by the chassis. In addition, the performance, latency, and jitter of the port itself are all factors in the performance of the switch.

David Newman, president of Network Test and member of the Network World Test alliance, said: "Everyone mentioned the line rate per port. Now the problem is how many ports can you reach the wire speed before packet loss occurs ." Based on published Device specifications, the largest enterprise-level Core switches include Force10 E1200, Foundry RX series, Cisco Catalyst 6500, and Extreme BlackDiamond. When comparing these specifications, Foundry's RX-16 switch is the highest-capacity switch that can run 0.64 million M Ethernet ports at full speed, in case of over-configuration, that is, the total bandwidth on all ports exceeds the capacity of the switch, a total of up to 1.92 million MB of ports are deployed. The Force10 E1200 TeraScale switch can run ports of 0.56 million MB or over 2.24 million MB. The Extreme BlackDiamond 10808 chassis supports 48 non-blocking 10-Gigabit ports. Cisco's Catalyst 6513 can process 32 10-Gigabit Ethernet connections in full duplex operation.

Some people think that, compared with how the vendor divides the bandwidth of each slot and the total system, it is not very important for the switch to handle variables such as Jitter and packet loss rate during full-speed operation. "I think latency is a more useful indicator than throughput," Newman said. Cisco is obviously the best in this regard ." He has measured a Cisco Catalyst 4948 device with a latency of about 3 microseconds at a 10-Gigabit speed. "This is the lowest latency in the device I have measured," Newman said. The latency of Force10 is usually several hundred times higher than that of a Cisco switch, which means that thousands of packets are not processed. But they have gradually solved this problem ."

Lawrence Berkeley National Laboratory LBNL, a Department of Energy Research Laboratory in Berkeley, Calif., uses both Force10 and CISCO switches at the data center and LAN core. Mike Bennett, Senior Network Engineer at the LBLnet Service Department of the lab, said it was very difficult to specify which of the two products was "faster" or "better performance" because these switches were in different applications. "I tested the 20 thousand and E1200 vswitches with port 6500 and port, neither overload nor packet loss," said Bennett. Therefore, it is not easy to say that a vswitch is faster than the other one, and they must run as effectively as described in the advertisement ."

Research on next-generation Architecture

At Lawrence Livermore National Laboratory at the other end of Berkeley, the network team is studying the next-generation switch architecture that enables Ethernet to reach the bandwidth and latency levels in the storage area network.

Dave Wiltzius, head of the lab's network department, said: "The performance of all next-generation devices will increase by an order of magnitude, and everything will be 10 Gigabit. Therefore, we need to find a switch or switch structure that can provide more than 2000 10-ge ports. In fact, we are interested in using a fat tree topology and something like this to create a converged switch environment ."

Wiltzius's fat tree topology involves a Mesh network) non-blocking switching architecture. This architecture is a bit like a traditional public telephone network, where switches are simple devices that are interconnected through multiple paths and have several connection ports. It is important that they can effectively use the bandwidth. However, Ethernet has not been able to do this, or it is not doing very well.

Wiltzius is using Port Clustering or layer-2 "hash" technology to achieve this fat tree effect, that is, bind multiple Gigabit or 10-Gigabit Ethernet links to a larger virtual pipe. Connecting a vswitch with a hashed Ethernet pipe or connecting the server to a vswitch can increase the virtual throughput, however, this connection is limited to eight ports, that is, a 10-Gigabit link with eight hashes can provide a maximum throughput of 80 Gbps ). This method uses an algorithm for randomly transmitting data packets on a bundled connection. Due to the random nature of the algorithm, the use of Hash technology will produce an uneven distribution, and may not necessarily provide the best bandwidth utilization.

Breakthroughs in data center Switches

Both Arastra and Woven Systems are emerging vendors actively seeking to achieve breakthroughs in exchange technologies. Arastra is a privately held vro company in Palo Alto. Woven Systems is a company in Santa Clara that develops Ethernet Mesh-based products in a semi-secret state.

"What we are doing is to provide the best fiber channel or InfiniBand features in a 10-Gigabit Ethernet architecture," said Harry Quackenboss, president and CEO of Woven ." The approach adopted by Woven is similar to grid or distributed cluster computing technology. In this technology, symmetric multi-processor SMP servers are replaced by single/Dual-processor nodes connected through the network.

Quackenboss said: "The same thing will happen in the LAN exchange technology of the data center. Large LAN switches are expensive. The biggest non-blocking switch you can buy for data center applications is the 64-port Foundry system ."

Woven is developing two-layer Gigabit and 10-Gigabit Ethernet data center switches that use special algorithms to simulate InfiniBand or fiber channel networks in some way. In this structure, multiple paths can be created between vswitches. Because the transmission channels are not disabled, such as spanning tree-based Ethernet), bandwidth can be allocated to these paths more dynamically.

Quackenboss said: "If you want to build a network composed of more than two switches, you can use the link aggregation or relay technology to bind the Ethernet network segments together. However, if you want to put three or more switches in the network, one of them will become a bottleneck ." Layer-3 switching technology, OSPF Open Shortest Path First, ECMP equivalent multi-path routing, and other protocols can be used to create a multi-path network, but these methods also increase costs. According to the IDC survey, the average cost of layer-3 switch ports is five times that of layer-2 ports.

In addition, changing the data center connection to multi-path L2 Ethernet can help administrators dynamically reconfigure applications and servers without actually recalling them. This idea is somewhat similar to the two-layer Metro Ethernet technology being developed by an operating device manufacturer.

Use Mesh to virtualize large Switches

Another company that has launched the product is Raptor Networks, which has a three-year history. The company produces fixed configuration Gigabit and 10-Gigabit Ethernet switches that are connected together to form a Mesh structure. Raptor does not focus on high-density data centers, but focuses on the aggregation of transmission streams between the local backbone network and the wiring room. Tom Wittenschlager, CEO of the company, said: "We have created capabilities to provide special features on the second layer, which can only be provided by other companies on the third layer ."