Friday, September 21, 2012

Cisco Nexus 3548 and Arista 7150: Duelling ultra-low-latency switches

 
 
Cisco and Arista Networks Inc. traded punches this week, both announcing new, ultra-low-latency top-of-rack switches that not only are fast, but pack many more features and functions than a typical switch in this class.
 
The Arista 7150 is the first switch in the industry to use Intel Corp.'s new Fulcrum Alta FM6000 networking chip, while Cisco's Nexus 3548 uses a new Cisco custom application-specific integrated circuit (ASIC), the Algorithm Boost or Algo Boost chip. At first glance, Cisco has taken a lead in the race to near-zero latency.
 

Arista 7150 and Cisco Nexus 3548: Fast and smart

 
Both top-of-rack switches push the state of the art in low-latency forwarding, a feature that is critical to the competitive high-frequency trading market and also attractive to high-performance computing shops, particularly in genomic research and oil and gas exploration.
 
The Arista 7150 has 350-nanosecond forwarding latency, a 30% improvement on previous generations of Arista switches. The Nexus 3548 ships with 250-nanosecond latency, a significant leap over Arista. The Algo Boost ASIC on the Nexus 3548 can operate in "warp" mode to push latency down to 190 nanoseconds. It achieves this by reducing the size of the switch's address table to from 64,000 to 8,000 hosts.
 
But these ultra-low-latency switches are smart as well as fast. They offer low-latency multicast and unicast routing and in-hardware network address translation (NAT).
 
The Arista 7150 now has "all the features and functions of a [Cisco] Catalyst 6500," according to Arista customer John Koehl, head of infrastructure and operations for Headland Technologies LLC, an algorithmic financial trading firm based in San Francisco and Chicago. As many financial trading firms do, Koehl collocates his switches with financial exchanges across the world so that trades can be made as close to the exchange as possible. In those environments, NAT becomes critical.
 
"If you're connecting to the exchange, [NAT] provides a little bit of security because you can mask what you're coming in as," Koehl said. "Second, you don't have to use up all the exchange IP addresses. You can use your private IP inside and just NAT to what the exchange is providing you."
 
Having features like NAT in an ultra-low-latency switch like the Nexus 3548 and the Arista 7150 means that network engineers don't need to place a firewall or another device inline to perform these functions in ultra-fast switching environments.
 
"We try to keep our switch footprint as small as possible so that we don't have a lot of switch hops," Koehl said. "So, the important thing at each data center is to be fast on the network and at the same time have enough switch port capacity and provide all the features and functionality that we need so that we ... can do everything in one box."
 
The Nexus 3548 ships with 48x10 Gigabit Ethernet ports and an 64,000-host address table and 16,000 IP routes. The Arista 7150 ships in three models ranging from 24x10 GbE to 64x10 GbE ports. It offers a 64,000-hosts table and 84,000 IP routes.
 

Arista 7150: Programmable forwarding plane for SDN and network virtualization flexibility

 
The Fulcrum Alta chip on the Arista 7150 has a programmable forwarding plane. Combined with EOS, Arista's fully programmable operating system, this chip allows the Arista 7150 to be upgraded to support new protocols in hardware without a device refresh. The switch is shipping with silicon support forVirtual Extensible VLAN (VXLAN), for instance, but Arista could easily add native hardware support for Microsoft's alternative Network Virtualization using Generic Routing Encapsulation, or NVGRE, standard with a simple software update. Arista demonstrated the VXLAN support at VMworld last month, showing how an Arista 7150 can serve as a gateway for attaching non-VXLAN network services to a VXLAN network.
 
As other protocols emerge in the software-defined networking (SDN) industry, enterprises will be able to migrate to those new technologies without ripping out hardware, according to Martin Hull, senior product manager at Arista. "A fully programmable software stack doesn't necessarily get line-rate performance in hardware. A traditional fixed-logic switch gives performance but doesn't give you flexibility. The solution to all this is a hardware approach that has flexibility in the data plane and a programmable software stack, which is where Arista sits," he said.
 
That means that the Arista 7150 will have the flexibility to change on the fly as applications come out, especially on the forwarding plane with SDN or if network virtualization techniques are being deployed, said Rohit Mehra, director of enterprise communications infrastructure for Framingham, Mass.-based research firm IDC. "Arista will be able to leverage these technologies and make changes without rewriting and reworking the ASICs. Redoing the ASIC can take anywhere from 12 to 24 months," he said.
 

Advanced analytics in an ultra-low-latency switch

 
Both the Arista 7150 and the Nexus 3548 ship with advanced analytical capabilities that can track and analyze latency spikes and buffer utilization. These capabilities allow enterprises to tune their networks to avoid microbursts that can affect ultrafast applications.
 
Arista offers a Latency and Application Analysis Package (LANZ), which provides detailed visibility into buffer utilization, and captures data held in the buffer when the switch experiences congestion. Arista also has added time-stamping to LANZ so that enterprises can know exactly when a latency microburst occurred. The Algo Boost ASIC on the Nexus 3548 has a similar analytics package that can operate in real time.
 
"We put functionality in the hardware to do fine-grained polling per port on buffer utilization," said Paul Perez, vice president and chief technology officer for Cisco's data center group. "Even down to the granularity of 10 nanoseconds, we can collect buffer utilization and extract that out into a software interface. It can be used in offline mode to do trend analysis, but also in real-time mode to be able to tune your environment."
 
Nexus 3548: Cisco's retreat from merchant silicon?
Cisco's first generation of ultra-low-latency switches, the Nexus 3000 series, used merchant silicon from Broadcom Corp. Rather than ride merchant silicon to even lower latency, Cisco elected to build the Algo Boost ASIC. Although Cisco has emphasized ASICs as a major differentiator, the company will continue to use merchant silicon where appropriate.
 
Cisco CTO Perez said his company believes in using custom ASICs "when you need it and merchant when you don't." "We have a non-religious technology strategy," he said. “We have 600 silicon designers and more than twice that in software developers to drive our custom capabilities. But we will take advantage of commercial silicon where appropriate."
 
Cisco will use much of the technology in the new Algo Boost ASIC to enhance other silicon in Cisco's product portfolio and not just in switching. "If you look in our server line, Unified Computing, we're differentiating and adding value in a highly competitive environment with custom silicon for expanded memory and also with our custom NIC[network interface card]," Perez said. "I think some synergy between this Algo Boost technology at the switching level coupled to that computing edge at the NIC is a very fertile area of exploration for my team in terms of how we can progress the engineering of the next generation of high-performance computing environments."
 

No comments:

Post a Comment