Mellanox Technologies just announced a couple of “data center connect” systems designed to allow companies to use fiber to link InfiniBand switches to create linked “cloud” data centers at distances of up to about 50 miles. This development makes what I think are two key points about cloud computing; first, we need to think outside the data center when we talk about it, and second that there’s been a switch technology (InfiniBand) available to support it for some time now, and it’s not set the world on fire.
Let’s start with the second point first. InfiniBand has exactly the same mission as a general “fabric” switch. Companies like Mellanox have been supporting it for over a decade, and probably many (most) of my readers have never heard of them. They’re arguably the market leader in InfiniBand, which is the most-deployed fabric technology, and they’re a shadow in the networking market. Which proves that if you look at data center network evolution in the future as being an extension of the trends of the past, fabrics aren’t going to mean squat. But they will, so you have to look at something more revolutionary than evolutionary, which brings us to the cloud.
Cloud computing is a resource-pool architecture, one that’s designed to break down the hard association between application and hardware that exists today even in nearly all virtualized data centers. Cloud computing is also an architecture designed to alter the relationship between workers and resources, to permit the closer coupling of compute-based tools to activity (what I’ve been calling “point-of-activity intelligence”. This new-style productivity support creates new benefits, which drives new spending, and it also creates more horizontal traffic. The cloud is also the model that’s emerging for the service logic of the next-gen network. Carrier interest in SDN and NFV prove that they’re interested in a cloud-hosted model for all of the network intelligence above simple forwarding.
Cloud computing as a service, if it fulfills its promise, would have enough demand to justify deployment in all of the metro areas with reasonable opportunity. Unless we presume zero competition we need to presume competitive overbuild too. Furthermore, if we assume that SDN and NFV go forward as operators want them to, we’ll be deploying hosted service-layer functions throughout not only these cloud-opportunity metro areas but the remaining ones as well. All of these service-layer clouds will have to be made up of distributed “data centers” that are likely located in all the current central offices and service points of presence. That’s a lot of data centers.
Building a distribute virtual cloud from discrete data centers is something that’s likely to require very high connect bandwidth. In fact, it’s very possible that the most important driver in designing a metro network will be intra-cloud traffic because this traffic will require QoS control, which best-efforts Internet service does not. The data center switch in this kind of picture isn’t just a data center switch, it’s an element in a metro-cloud matrix, and it has to be architected for that purpose.
We see, in today’s data center networking market, a division. Some vendors, including Mellanox Technologies, have a fabric but not a metro-cloud fabric because they don’t have the broader positioning to support that larger mission. Some vendors, including Cisco and Juniper, have a metro-cloud vision that’s more inherited from the fact that their switching implicitly supports interconnect than that they’ve really designed themselves for the mission. A very few (one that we know of, Plexxi) has an architecture that’s arguably designed for the metro-cloud mission. We think Juniper is probably (belatedly) pushing its QFabric toward that mission.
The mission of metro cloud is a mission of SDN and NFV, which means you have to be able to articulate an SDN position that’s complete in order to claim support. Right now, we don’t have any articulations that completely meet the test, as I’m finding as I develop our comprehensive view of SDN models, implementation, and market impact. What everyone who’s interested in SDN, NFV, next-gen network architectures, metro, or the cloud should be looking for in the coming months is a maturation of the metro-cloud vision of the vendors and their effectiveness in making their SDN strategy both a logical metro strategy and a logical cloud strategy. Will any meet the test? We can hope!