Very good questions – would love to get into details after we unveil our product later this year and discuss how we can independently scale data from compute for virtual machines. For now, I’ll drop a couple of hints in response to your questions:

1. Virtualization (hypervisor) provides us a nice container that can help compute and storage grow and shrink independently for VMs.
2. A single server-attached SSD can replace hundreds, if not thousands, of disk drives in the data center. Imagine what they can do in a scale-out cluster.

I do see value for companies which are into data crunching (more cpu’s and more data at same rate), not sure if it is for masses though.

Thanks for your comments on the appliance approach. We are focused on making the virtualized datacenters very easy to deploy and manage and the appliance experience is a part of it. The other key aspect is high availability. As you pointed out, the traditional architecture mandates the use of network storage to keep data safe even if compute nodes fail. Over the last decade or so, distributed computing software written at companies like Google, Yahoo, Amazon and now Microsoft (Windows Azure) has changed the game though. This kind of technology enables data centers that combine compute and storage in a scale-out architecture and are extremely fault tolerant to disk and server failures. Data is replicated across the cluster and is always fully protected. Our system architecture follows similar principles of distributed computing and provides advanced data protection capabilities in a converged architecture.

Ajeet Singh
Chief Products Officer
Nutanix

In addition to the issues that arise from dealing with these small details, is that usually systems like these are built and tuned with a particular architecture in mind (ratio of SSD to spinning disks, for example). With software-only solutions, customers can play with these settings and get caught in unexpected use cases for the software. Control the full-stack from software to hardware — and you know what customers are doing.

The challenge with blending compute & storage that I’d love to see an answer to is this: Compute and storage have different durability properties. It’s acceptable to occasionally lose a compute node/virtual machine/instance. It’s not acceptable to lose customer data. Does blending compute with storage add any additional risk in the complexity of the system that may put durability at risk?