10-21-2013, 09:59 PM
Well, you could either write your own software, or cluster the nodes.
If you choose a clustering solution, you can cluster multiple hosts that have X number of GPUs each.
e.g.:
Host1 has 2 GPU
Host2 has 3 GPU
Host 3 has 2 GPU
Run vcl or disthc on all 3 nodes and you can effectively use 7 GPUs in a single job.
Which solution you use depends largely on hoe you are going to use it and how much money you can afford to sink into it.
vcl is a great solution for high speed LANs (10 Gbps +), and disthc is pretty much everything else.
If you choose a clustering solution, you can cluster multiple hosts that have X number of GPUs each.
e.g.:
Host1 has 2 GPU
Host2 has 3 GPU
Host 3 has 2 GPU
Run vcl or disthc on all 3 nodes and you can effectively use 7 GPUs in a single job.
Which solution you use depends largely on hoe you are going to use it and how much money you can afford to sink into it.
vcl is a great solution for high speed LANs (10 Gbps +), and disthc is pretty much everything else.