IT infrastructure-as-a-service provider Savvis is one of the 10 organisations that beta tested Cisco's Unified Computing System datacentre platform. CTO Bryan Doerr took some time to share his thoughts on the UCS platform and next-generation datacentre architectures.
Can you share your initial thoughts on the UCS platform?
The basic capabilities they advertised, they're there. It continues to evolve in terms of the [user interface], the polish. The first [testing] steps are the interfaces; then we'll dive into scale testing, embedded functions, things built for virtualisation -- the primary benefit of this box -- as opposed to being built physically and then virtualised.
But this is the culmination of a vision-sharing session that took place a couple of years ago. Cisco and Savvis realized we had very like visions.
How long will testing take before you make a deployment decision?
We're a good part of 2009 away from a purchase decision.
What other vendors' data center unification solutions are you evaluating?
One of the problems is we don't have that many choices. That's what the novelty is here. [Cisco has] forced the industry to think of another evolutionary path for the server. We were having trouble motivating that original thought and revisiting the current mode. It's such a critical component of the datacentre and one [where] there's so much allegiance to a particular style of packaging and tool sets that introducing a brand-new change was going to be pretty difficult.
At a high-level, what they've done is reintroduced the idea of the server and forced people to reckon with whether or not this style and subsequent need for virtualisation capabilities are key to enabling a different style of service. We have to continue to look at that solution versus traditional solutions. There are clearly advantages. Where this box has to come out and shine is in operational efficiencies; it has to drive cost down, [perform] datacentre consolidation in terms of fabric and network interfaces into a cost reduced connectivity strategy, [and derive] related skill savings in terms of people that run [storage-area networks] versus people that run networks — all that segregation of skill and technology needs to produce operational benefit that we can quantify. That's what we're going to have to consider against entrenched solutions.
I intuitively believe that for service providers like Savvis to obtain the differentiation of service level that's going to [drive] a substantial move towards managed services; we're going to have to show a level of cost benefit, functional benefit, that most enterprises find out of reach. It's too easy to continue in the mode of buying and using the traditional approaches -- it's going to be when they see a cost benefit, a control benefit, a feature benefit that they can't obtain themselves at a comparable price point, that enterprises are going to re-envision how they deliver infrastructure. These whole system solutions are a piece of that story.
I also think these systems are necessary to engineer the quality-of-service levels in multitenant environments. If you don't have the deep QoS controls then at some level you're sharing assets. Enterprises will frequently use the risk of sharing -- or the failure to get resources when needed and the adverse impact on applications -- to justify dedicated deployments in the traditional mode. We need to move away from that for service providers to really show their stripes.
How much involvement did Savvis have in the development of this product?
The way it's built and the way it scales is such that it gets more affordable and beneficial as you buy in significant quantities -- as you operate large numbers of servers. You can attach behaviors with virtual ports that move with virtual machines, and the way that profiles follow server definitions through a logical configuration that rests over the hardware environment -- those are features that enable efficiency when you're dealing with lots of customers, particularly in a multitenant environment.
We met with Cisco to review features sets, provide comments on the relative importance of features. There was input from Savvis on multiple occasions. What brought my involvement up was actually a conversation I had with some Cisco executives back close to two years, when I laid out this vision for an entirely virtual application infrastructure starting from the network endpoint all the way through the wide area and down to the data center, preserving QoS through that entire path but having no equipment dedicated to any particular customer. We needed to get more of that functionality on the server/switching environment.
Did your traditional server vendors react to this with solutions of their own? (Savvis' blade server vendors are HP and Egenera)
I haven't been approached yet by folks who want to argue for an alternative solution. But there's no question that other [vendors] are capable and have the technical ability to do the kinds of things Cisco has done here. But the question is, where is the will and the motivation? Do they see the market that Cisco sees and do they see a risk to the product that they already offer in that new market opportunity? It's one thing to say you can accomplish all the functions of this unit by assembling these four boxes and layering this hunk of software on top; it's another to buy the unit with all of that done. It's a fully integrated system, it's a new architecture for delivering the functions that look familiar. You can't argue equivalence by finding three or four other boxes that you can integrate together and claim the same function. The traditional [vendors] now need to take a look at this again, afresh and say, 'do we believe there's a market for this platform built for virtualisation; or do we feel that the current evolutionary path of the server in its current form is the one that's likely to win?' I think their investment will follow that decision.
Keep in mind Cisco wasn't the first company to envision a capability like this. Savvis has used Egenera, which has had stateless nodes in an integrated backplane for a number of years. We used them in our utility platform, quite effectively. Cisco has done things that are new and different, more scalable and several ways better; but there's an ancestor to this architecture and it's been around for several years.
What specific pain points are you looking for UCS to address?
One of the things we're interested in doing is getting more virtualisation for our spend. We see an active approach to dealing with memory constraints and the creation of VMs, and the fact that you get more scale out of it. We are interested in a better understanding of the service profiles built into the environment, and in a multitenant environment in particular, and having those profiles follow the virtual machines themselves as opposed to needing to relax much of the profile definition in the virtual world, just so that as VMs migrate they're compatible wherever they land. So instead of running wide open we can run more locked down and have migrations trigger profile movements as opposed to not. The effect of that is that it starts to make the architecture more secure. That's one benefit that we're looking for.
Running traditional server architectures today where you have dual HPA and dual [network interface cards] and SAN fabrics and switches, and Ethernet fabrics and switches -- all of that redundancy and our inability to tailor the utilisation at the network level through those paths on a per-customer basis are all limitations to our ability to provide that truly virtual infrastructure. And their unit, presumably, solves many of those problems.
By being an integrated unit, we have one interface point that we can configure network connectivity through, and all kinds of network connectivity and server configs, so it simplifies the complexity of integrating new technology to our backbone management system. We're looking for those kinds of features to demonstrate themselves in the evaluation.
Are you also evaluating Juniper's Project Stratus?
I am not at this point. I don't have any immediate plans but I think that anytime we're getting ready to do a study with the kind of impact this selection would have, one of the things we have to do is look across the entire industry and make the best choice that we can. It would be important for us to review everything in the marketplace in that decision.
Our study for selecting next-generation hardware hasn't formally commenced. Once that's defined we'll canvass the people we want to take a look at.