Blog Post

Optimizing the Virtual Data Center

The promise of virtual machines is that operators don’t need to worry about where their servers are. You can have one big server running on five physical computers, or a hundred tiny servers running on one physical machine. This makes it easy to adjust capacity; it also means creating a new server is as simple as dragging and dropping.

But while data center operators might not care where their servers are, the servers do. Today’s data centers are based on Web Services and SOA architectures. Instead of one big mainframe, we have many small servers all talking to one another.

In a traditional data center, machine-to-machine conversations like these can take milliseconds, resulting in slow applications. But if chatty virtual servers live on the same physical machine, they can communicate in microseconds.

Done right, putting virtual servers that need to talk together on the same physical machine could make applications a thousand times faster.

What we need is software that continuously analyzes conversations between all servers, then automatically reconfigures the data center so servers that communicate more often are on the same physical hardware. Call it a Virtual Data Center Optimizer.

The optimizer would be constantly moving servers among physical machines to find the optimal configuration as the data center changes, much as the Internet’s routers are always looking for the best path between two points as the network changes.

Companies already make tools to help humans find good configurations and to manage the flood of new virtual machines. Cirba, for example, helps administrators decide where virtual machines should go based on policy and workload. Embotics, on the other hand, tracks the lifecycle of virtual machines to address sprawl and security challenges.

Firms like these, along with technologies like VMWare’s VMotion, are well-positioned to optimize server-to-server communications. But we need more than just a management tool — we need an autonomic process that’s constantly adapting. Whoever solves this problem stands to speed up data centers dramatically.

8 Responses to “Optimizing the Virtual Data Center”

  1. Actually, the seed of this post was an amazing presentation Steve Shah of Risingedge ( gave at Interop last year ( on “the state of the cage.” He looked at several trends that were converging, and concluded (among other things) that the performance and efficiency issues were moving into the data center fairly quickly.

    Not sure if Steve still has the material but it was awesome (no surprise as he’s a superlative presenter and great thinker.) He’s now a founder of stealthy startup Asyncast ( but he won’t tell me what they’re up to yet.

  2. Ken,

    Agreed, there is numerous “costs” that need to be taken into account, including the ones you mention, some real, some opportunity or technical costs, costs of migrating between servers vs network overhead vs the benefits that John mentions etc.

    We’ve been doing a lot of work on this as part of the future development of our computing platform FlexiScale, and hope to show some of the first examples of this part of our technology later this year but it’s definately an area with a lot going for it at the moment.

  3. John Gannon

    There are some security and fault-tolerance benefits that come from having VMs that communicate with one another on separate hosts. This should be weighed against any reduction of latency associated with cutting out external network communications.

  4. The concept of a data center “optimizer” is dead-on, but the analysis is a bit more advanced. There are actually lots of variables to take into account, not just proximity of processes and the pipes between them. Consider you need to know the workloads, the capacities of different (physical) machines, the cost to operate each machine (capital cost, power cost, etc.) the desired networks, the compatibilities of machines vs. software, etc. and optimize against these in real time. Add in the fact that some applications may be virtualized (on different vendor VMs), and some may be native — and in fact, some may be “scaled-out” across a server farm.

    The good news is the concept of policy-based optimization is very real. (BTW, Gartner calls it “Real-Time Infrastructure” while Forrester calls it “Organic IT”; others call it automated workload management). At any rate, this sort of optimization can both speed-up applications, and reduce operating & capital costs by 1/3 to 1/2. Check out companies like