Separate vs. Combined Server Clusters for App Workloads and Shared Storage

A widely used "separate-cluster" architecture for datacenters uses a server cluster, sometimes called a "storage array," that implements highly available shared-storage volumes, and a separate server cluster running application workloads that access those volumes.

An alternative "combined-cluster" architecture, sometimes called "hyperconverged," uses a single server cluster in which every node can participate in implementing highly available shared-storage volumes, and also run application workloads that access those volumes.

For each of these architectures, there are many commercially available implementations. Using technical (not marketing) language, and without naming specific products, this tutorial evaluates key strengths and weaknesses of each approach, including some practical issues that are often omitted in such evaluations.

Craig Dunwoody is co-founder and CTO of two Silicon Valley companies: GraphStream, an integrator of advanced scalable data infrastructure, and Birchbridge, an early-stage startup that is developing a new cabinet-scale datacenter building block product with an innovative physical-layer architecture. Previously, at Silicon Graphics, he developed system software for seven successive generations of industry-leading visual computing systems. He earned BSEE, MSEE, and MSCS degrees from Stanford University, and has co-authored five issued and seven pending U.S. patents.