Interop Insider #4 (MP3): Arista Networks Positions Itself As "The" 10-Gb Networker Of The Clouds
Although she dodged the question at the end of my podcast interview (below), Arista Networks CEO Jayshree Ullal can't help but think that history is going to repeat itself. Following Cisco's acquisition of Crescendo Networks in the '90s (where she worked), Ullal ended up working for Cisco for 15 years. Notwithstanding its newly announced Intel-based blade servers, Cisco usually prefers to buy than build. If everything Ullal says about Arista's 10 Gigabit Ethernet switches and "the cloud's" appet
Although she dodged the question at the end of my podcast interview (below), Arista Networks CEO Jayshree Ullal can't help but think that history is going to repeat itself. Following Cisco's acquisition of Crescendo Networks in the '90s (where she worked), Ullal ended up working for Cisco for 15 years. Notwithstanding its newly announced Intel-based blade servers, Cisco usually prefers to buy than build. If everything Ullal says about Arista's 10 Gigabit Ethernet switches and "the cloud's" appetite for them is true, she'll probably end up at Cisco again (by way of acquisition).
If Cisco doesn't end up acquiring Arista, maybe Sun will. After co-founding Sun Microsystems (his claim to fame is being employee No. 1) back in the 80's, Arista co-founder Andy Bechtolsheim has busied himself with building and selling companies to -- you guessed it -- Cisco and Sun. Bechtolsheim and Ullal worked together while both were at Cisco and now the dream team is back together at Arista expanding on their high-speed networking pedigree.
Click the tiny play button to listen to to my podcast interview with Jayshree Ullal.
For their encore performance, Ullal, Bechtolsheim, and the other co-founder David Cheriton are once again going after "purpose-built" (as opposed to mainstream) gear. In this case, Arista believes it's on to something with its 10-Gb Ethernet switches, particularly when it comes to high performance and cloud computing situations where a lot of data is getting moved around and there's little tolerance for latency or downtime or the traditional expense associated with overcoming either problem in mission critical environments. For example, the trading floor of a stock exchange where availability is imperative and nanoseconds can translate into millions of dollars (giving new meaning to "time is money").
The inspiration for Arista's 7100 series switches, says Ullal, comes from the efficiencies that are characteristic of cloud computing. On average, typical IT infrastructures -- the systems, the storage, and the networks -- run below 50% utilization (Ullal cites 20%). But one of the driving forces behind cloud computing is virtualization and the push towards using every single cycle [sic] of compute, storage, and networking power. With the processing power of modern servers doubling (or more, thanks to multicore) every 18 months (Moore's Law) and those servers (where deployed in clouds) running at near 100% utilization, the network becomes the bottleneck and it's those environments that Arista is specifically looking to target.
Ullal claims that Arista's 7100 series of 10-Gg switches' port-to-port latency comes in at around 600 nanoseconds -- performance that she claims only InfiniBand comes close to meeting. But what could really make Arista's gear interesting to the operators of public and private clouds is the operating system inside the switch (Arista calls this the Extensible Operating System or "EOS"). According to Ullal, regardless of what agents are running in the switches (routing protocols, command line interfaces, etc.), the switch can be live-patched without having to bring down the network it supports.
Part of the EOS, according to Ullal, is a database to which the switch's state is constantly mirrored. Should a patch have to be deployed to the switch, the network barely hiccups as the state information from just before a patch's deployment is recalled from the EOS's database.
For networks that can't tolerate any downtime, availability is often guaranteed through the deployment of redundant switches. When one switch is brought down, a standby switch takes over. The problem with this architecture is that the cost per supported device (server NIC or storage) link doubles. Doubling or tripling up on "older" 1 G-bit switched Ethernet links for the sake of performance can also result in the same net effect on cost (double or triple the cost). At $500 per 600 ns latency port without the need for redundant switches (because of the live-patching capability), Ullal claims there's no match for Arista's switches (which start in the $10,000 - $20,000 range).
If that's truly the case (admittedly, I don't have all-seeing visibility into the market for 10-Gb Ethernet switches), it's only a matter of time before this networking team cashes in one more time. Give the podcast a listen and tell me (and the rest of TechWeb/InformationWeek's audience) what you think in the comments are below.
David Berlind is an editor-at-large with InformationWeek. David likes to write about emerging tech, new and social media, mobile tech, and things that go wrong. He can be reached at [email protected] and you also can find him on Twitter and other social networks (see the list below). David doesn't own any tech stocks. But, if he did, he'd probably buy some Salesforce.com and Amazon, given his belief in the principles of cloud computing and his hope that the stock market can't get much worse. Also, if you're an out-of-work IT professional or someone involved in the business of compliance, he wants to hear from you.
Twitter: (@dberlind)
My Facebook Page (Facebook should have a namespace like Twitter, FriendFeed, and the others)
Flickr (davidberlind)
YouTube (TechWebTV)
FriendFeed (davidberlind)
Del.icio.us (dberlind )
Me on LinkedIn (LinkedIn should have a namespace as well)
Plaxo (davidberlind)
Disqus (DavidBerlind)
About the Author
You May Also Like