I am trying to build the cuda cluster with the single node for testing the certain issues, I googled it to find some hints on it, but cannot find out the reliable solution .
How can I really create the GPU cluster with a head node and a single node???
You mean something like this?
[url]http://devblogs.nvidia.com/parallelforall/how-build-gpu-accelerated-research-cluster/[/url]
It would seem like you’d need some Infiniband cards to accomplish that, but I’m not familiar with those, perhaps someone else can comment.
It seems the graph is rather conflicting, as it has infiniband interconnecting only the compute nodes, but in fact in the text it mentions installing infiniband drivers on the head node… which sounds correct. From what I can tell, the difference in the cheap vs expensive infiniband cards is number of ports and bandwidth quantity that can be carried per card/port.