Start-up Woven Systems this week is expected to launch its 10G Ethernet data centre switch -- a box the company says one-ups the switching power of Cisco, Force10 Networks and Foundy Network 10G gear and matches the low-latency performance of Fibre Channel and Infiniband with lower-cost Ethernet technology.
At 144 non-blocking 10G Ethernet ports, Woven says its EFX 1000 switch provides the highest port density and capacity in the industry. The switch includes low-latency packet processors and congestion-detection mechanisms, which mimic Fibre Channel or Infiniband performance but with less cost and better manageability, Woven says. As many as 4,000 of the 10G switch ports can be configured as a single virtual switch, or fabric, for connecting servers, storage and other network devices in a data centre at very high speeds.
The 12-slot EFX 1000 chassis uses 12-port 10GBase-CX4 Ethernet line cards and has a redundant backplane with a capacity of 2.8Tbps to allow every port on the switch to run at line-rate. Speeds and feeds aside, Woven says the key to its product is its vSCALE packet processors -- three on each blade -- which allow the boxes to connect as a fully redundant, multipath mesh while using standard Layer 2 Ethernet technology.
This setup differs from standard Layer 2 networks, where Spanning Tree Protocol closes off all but the primary uplink path of a LAN switch, only activating the alternate paths when the links fails -- a process that can take up to a half-second or more.
Woven Systems says its technology allows a switch to have multiple paths to other switches in the network that are always active and can move traffic among any other switches in the mesh. The multipath Layer 2 network also allows for near-instantaneous failover of network connections to alternate paths, with switchover time in the sub-10-microsecond range (10Â millionths of a second). Traffic latency from port-to-port inside a system measures around 1.5 microseconds, and latency among interconnected Woven switches is around 4 microseconds. The company says this low-latency performance of its switches is on par with other high-speed/low-latency interconnect technologies, such as Infiniband, Myrinet or Fibre Channel.
The vSCALE packet processors analyze Layer 4 traffic flow -- TCP/UDP port data in packets --- and measure traffic congestion and link integrity based on the flows they observe. If the switch detects high utilization on one uplink port, the switch moves the flow to another live port. Switches connected in this fabric insert traffic measurement bits into every Ethernet frame moved by the Woven switches. (The time-stamped data goes into the extra space of the virtual LAN [VLAN] tag frame space). For every hundredth Ethernet frame transmitted, this time-stamped data is pulled from a random packet and sent back to downstream switches to give the devices data about latency and congestion on the network. The downstream switches use this information to make Layer 4 load-balancing decisions across the multiple paths to which they are connected.
This approach "is an adaptation of flow-based traffic forwarding," says Joe Skorupa, a research vice president at Gartner. "Similar to Layer 3 networks, but they're doing it in a very innovative way down at the Layer 2 level." Layer 2 multipath switching allows for much quicker packet processing than Layer 3 routing protocols, such as OSPF, EIGRP or RIP.
"Doing all the switching at L2, [Woven] can use commodity Ethernet switch chips," Skorupa says, which make the products less expensive than Layer 3 gear.
To create the multilink mesh, the Woven vSCALE processors play a trick on standard Layer 2 VLAN, spanning tree and MAC addressing technologies. The Woven switches establish multiple Layer 2 VLAN connections among other switches in the mesh. For every link to another switch in the mesh, media access control (MAC)Â addresses of Ethernet frames on the switch (which are associated with the VLANs) are encapsulated inside other Ethernet MAC addresses. This method invokes multiple instances of Spanning Tree (one instance per encapsulated MAC/VLAN association). The result is that switches can send traffic down various multiple paths very quickly, simply by changing a bit in the MAC/VLAN association in frame headers.