We are looking into network switches able to support an HPC Cluster of up to 32 nodes, but probably closer to 10/15. There will be one login and storage node, and the rest will be computing nodes.
We will probably tend, software-wise, towards multithreading rather than distributed computing, so there's shouldn't be a lot of inter-node communication, although the data would be on a network filesystem so I expect a lot of network traffic between the login/storage and the computing nodes.
The Nexus 3132C-Z is still affordable for us, and has enough 100 GE ports for our needs, and a 6.4 Tbps capacity.
What worries us is the buffer size, only 32MB. Would this be a bottleneck in the above scenario?
Alternatives like the 3636C-R are a bit too expensive.