Networking
Compute and storage can be sized correctly and the network can still become the bottleneck. I look at how bandwidth, NIC choices, fabric and topology line up with the rest of the configuration.
What I work on
Making the network part of a validated configuration rather than a line item added at quote time.
What I look at
Whether the NIC and port choices in a configuration can carry what the compute and storage sizing assumed. Whether the topology a customer already runs supports the design. Whether traffic between servers has been considered, or only the link out of the rack.
What changes a configuration
Adding accelerators usually changes the network as well as the power budget. Moving from local drives to a shared storage platform moves traffic onto the fabric. Both are decisions made somewhere else that land here.
Getting it wrong usually shows up late, as a deployment that misses the performance the proposal implied, and that is expensive to fix once everything is installed.
Where the network decides whether training scales
kushagragoyal.com/stack/networking