IBM Power Support in South Korea After End of Life

The Nodes Do Not Own Separate Storage

In a parallel filesystem the nodes serve pieces of one namespace to each other rather than holding independent volumes. That is what delivers the aggregate bandwidth the design exists for, and it means the boundary between a local failure and a cluster failure is much weaker than in a conventional estate.

A Local Failure With a Distributed Effect

When a node loses its disks, part of the shared namespace goes with them. Every other node needing that data either waits or participates in reconstructing it, so a single-node hardware fault presents as the whole cluster slowing down. Replication keeps the data available and cannot make the reconstruction free.

Which Reprices Node-Level Storage Components

A drive in a parallel filesystem node is not equivalent to a drive in a standalone server, even when it is physically the same part. Its failure reaches every consumer of the filesystem, which makes it worth holding on a different basis than a general-purpose spare — priced against cluster-wide impact rather than one machine.

The Cluster Network Is the Other Half

Filesystem traffic between nodes runs over the interconnect, so an adapter or optic failure degrades the filesystem rather than merely reducing connectivity. ICD supplies node storage, storage adapters and high-bandwidth cluster networking for POWER8 and POWER9 generations past IBM end of support, shipped to South Korea in 4 to 7 days, covered for one to three years.