About the Event
In 2017, global data centers used more than 400 terawatts (4.16 x 1014 watts) (or about 3% of the total electricity), nearly 40% more than the entire United Kingdom. And this consumption at least doubles every four years. A large-scale data center consists of 100,000's of servers, and for some applications, a single service utilizes over 10,000 servers. Each of these servers produces and consumes 10's of gigabits per second of data traffic to carry out their distributed computations. To satisfy these growing needs, individual network link speeds have gone from 10Gbps to 100Gpbs, with 400Gbps links starting to emerge. Likewise, switching chips have gone from a capacity of 100's of Gbps to 6.4Tpbs, and extending beyond 12.8Tpbs. Until recently all of these networks were hardwired, implying that they only gained new capabilities at the rate of the development of new silicon.
In this talk, I will outline the overall state of our current high-performance data center networks, and then I will detail a new type of switch architecture—PISA, or Protocol Independent Switch Architecture, that enables programmable data center networks. Not only does this new architecture continue to scale the networks in throughput, it also significantly increases their capability at the rate of software development. In particular, I will discuss how PISA switches can be programmed to allow unprecedented insight into the flow and latency of packets in the network. I also will discuss new routing algorithms that can be tailored to a given computation and network load. Finally, I will give examples where the network fabric itself aids in the computation and synchronization of the distributed software, running not just on the servers, but within the interconnect itself. You will see how the emergence of PISA has greatly accelerated innovation in networking and distributed computation.
Dan is the Chief Development Officer and a co-founder of Barefoot Networks, a start-up that has created software and silicon, allowing unprecedented programmability and performance in data center networks. Previously, he was a VP of Engineering at Apple and a SVP at Cisco responsible for Cisco’s server and data center access switches. Dan joined Cisco through the acquisition of Nuova Systems in 2008. Before that, he was a VP at Google responsible for the platform team which built hardware and system software for Google's data centers. Dan was also a founder and VP of Engineering at Growth Networks, which was acquired by Cisco in 2000. And from 2000 to 2005, Dan ran the development for Cisco’s stackable switching products and Cisco’s core router, the CRS-1. Earlier, he held senior engineering and management positions designing computer systems for Silicon Graphics, Sun Microsystems, and Tandem Computers. Dan received his BSEE degree from Caltech, and his MS and PhD degree in EE from Stanford.
- Please RSVP by 5 pm on August 14th
- Those who have not registered will have to wait at the door as registered alumni are seated first
- Pay at the door. $20.00 including lunch, tax and gratuity. $5.00 for those not eating the buffet. (Please state "not eating" after your name when you RSVP)
- Please RSVP if you plan to attend. All registrants are committed to pay.
Questions? Contact: : Peter Tong (MS '81, PhD '85), email@example.com