The (net) world according to Andy Bechtolsheim

The (net) world according to Andy Bechtolsheim

SAN JOSE, Calif. – The technical horizon in networking as not as bright as it is for semiconductors, but it’s pretty darned good, according to network guru Andreas von Bechtolsheim.

Andy should know. He’s been designing high-end network switches since he founded Granite Systems in the go-go days of 1995. Cisco Systems bought the company for a cool $220 million, bankrolling this engineer’s engineer who passed on the favor when Google’s founders’ came calling. He wrote their first check for $100,000.

[Get a 10% discount on ARM TechCon 2012 conference passes by using promo code EDIT. Click here to learn about the show and register.]

The bad news is networking is not keeping up with the heady pace of Moore’s Law because of I/O bandwidth bottlenecks. Instead of doubling as chips do in transistor density every two years, networking bandwidths double about every four years (see below).


Click on image to enlarge.
Bechtolsheim gathered his thoughts in a keynote at the Linley Tech Processor Conference here.

“Moore’s Law is alive and well, and there has been nothing like it in the history of mankind,” he said. Thanks to solid work on serdes and merchant chips, networking is set to continue its traditional bandwidth “doubling every four years or so--it’s a nice improvement, but not at the rate of Moore’s Law,” said Bechtolsheim who briefly worked for Intel when he first arrived in Silicon Valley in 1977.

Bechtolsheim gets deeply technical at the drop of a slide rule. Discussing the impacts of data centers on network design, he quickly dropped down two levels to note a new DCTCP protocol can reduce by ten percent or so the need for mega-buffers on aggregation switches—an interesting detail.

But his bigger point was there is another wave yet to come. Network protocols and apps need to be tuned to better understand the underlying networks capacity. That’s heavy lifting, so in the meantime network switch design is all about right sizing buffers, he said.

Again without a breath he jumps down two levels in the technology to compute the requirements. “That’s [a buffer of] 100 Kbytes per flow and 10 flows per server, so a Mbyte per server and 10 Gbytes be switch,” he said.



Next: Too many choices
TAG:Andreas Bechtolsheim Software Defined Networking Silicon Photonics 100GE Cisco Luxtera Fulcrum Broadcom Networking Communications Switches ASICs OpenFlow