The big six cloud companies are on CAPEX growth trajectories that will take them past the traditional telcos in a few years, said Andy Bechtolsheim, Chairman of Arista Networks.
There is a race to build gigantic data centers that are more efficient and powerful than anything seen before. Speaking at the Open Server Summit in Santa Clara, California, Bechtolsheim said the expected gains from Moore's Law coupled with promising developments in silicon photonics and virtualization technologies, make it likely that the hyper-scale data centers will continue to hold a competitive advantage over the coming decade.
So, how do you build a network for a data center with 100,000+ servers and millions of VMs? Bechtolsheim said his company has already seen (and won) competitive bids for 100,000+ ports of non-blocking 10 GigE server connections. The ideal network, he said, should be truly transparent, flat and universal. This means the bandwidth and latency between any two servers in the data center should be the same. A top-of-rack switching architecture is preferred. To scale it for the biggest of data centers, Arista has developed a "Spine-of-Spine" network architecture that could scale to link up to 884,000 servers at 10G with a 3-1 oversubscription.