Hello,
in most threads about VLC/IB clusters, 4 x SDR hardware is recommended.
As I understand, latency is the most critical factor for a VLC/IB cluster.
What is the reason that 4 x SDR hardware is recommended? (Price? Well suppored by the VLS software)?
Thank you very much for any feedback!
John
By VLS I think you mean VCL. And no, Inifiniband is not supported by VCL at all. All VCL supports is TCP/IP. But you can do TCP/IP over Infiniband. So in that regard it doesn't matter what speed Infiniband you use.
4x SDR is only recommended because it's dirt cheap, and it is sufficient for hitting your target latencies. You can get all the 4x SDR hardware you need for just a few hundred dollars.
I appreciate the following update/response. I was looking into tackling Infiniband technology/configuration for the first time as well. But I was wondering how the Master node was going to communicate effeciently with the other systems. I did not want to invest in an additonal HCA in the master node if there was another possible configuration. Eg my current master node will not have any GPU's in the system; however it does have x4 ethernet ports. However the infiniband switch does not have any ethernet ports at all.
From what I see above and know now I should be able to configure another Ubuntu VM as the cluster master to configure all the jobs.
https://hashcat.net/wiki/doku.php?id=vcl_cluster_howto
The broker node is essentially the cluster master – it is the system that controls all the other nodes in the cluster. It does not need to have any GPU devices installed, and it should not have the AMD Catalyst driver installed. This node can be just about anything – even a virtual machine – as long as it has at least 8GB of RAM. It is the only node that you will interact with directly.
If you want to do Infiniband, then you have to invest in an additional HCA for the broker node -- that's the entire point. If your broker node is not using Infiniband, then the entire cluster is not using Infiniband.
Although I never was able to get it to work, the wiki talks about setting up the broker on a compute node as well. This could eliminate the need to have another server.