Hyper-Converged Scale Out vs. Scale Up

I received ample feedback from the vendor community regarding my Sydney, Australia VMUG Keynote. I discussed the considerations for when to select hyper-converged infrastructure (HCI). If you aren’t familiar with HCI, I wrote an introductory over on TechRepublic.com. One interesting point of contention is scale out vs. scale up. One of the primary design characteristics of HCI solutions is that most allow for natural “webscale” expansion via adding additional nodes to the cluster. 

When these solutions were first introduced, this meant adding both compute and storage capacity with each expansion. It provides a simple linear scale-out design that improves both compute and storage performance. The disadvantage is that it’s inefficient. Customers unnecessarily expand storage capacity when only needing compute. Most major HCI vendors alleviated this challenge by offering compute and storage nodes individually. Storage and compute now scale out independently. 

Notice I said storage and compute scale out instead of scaling up. The fundamental architecture of hardware based HCI solutions is a scale out design. Individual nodes are not intended to scale up. Meaning, if you need to increase the total memory of your cluster, you must scale out opposed to popping the cover and adding memory. Depending on the nature of your workloads, this could be a challenge. 

HCI hardware vendors have good reason to limit choice in scaling up post implementation. The advantage of HCI is the tight integration of the management and hardware stacks. Restricting the number of configurations reduces support issues. Software only vendors such as VMware with its VSAN solution don’t give you the integrated hardware management but do allow you to scale up individual subsystems.