Unsolved
This post is more than 5 years old
1 Message
0
984
July 20th, 2016 03:00
NANON/NENON VS ANON/ENON performance in mixed node cluster
Hi,
I was wondering what best practise are related to NFS networking for a cluster with 5 x X410 and 6 x NL410, where 500 NFS clients in a HPC cluster access the Isilon cluster concurently. Should the HPC cluster talk to all the 11 nodes or only the 5 X nodes? There is high plausability that the data accessed by the HPC cluster will reside on the X nodes, and I would assume the likelyhood of finding the data locally on the node a client are talking to are 1/5th if only talking to the X nodes, while 1/11th if talking to all nodes, and then have higher latency for accessing the data?
BR
Roger
No Events found!
sjones51
252 Posts
0
July 20th, 2016 11:00
Hi rkvam,
NANON is briefly discussed in the Isilon External Networking Best Practice Guide on page 9:
https://support.emc.com/docu58740
File pool policy will play a part in determining what is best for your cluster. I am working under the assumption that you have a pool to store the new data on the X nodes, and archived data on the NL nodes in some fashion.
The short answer is that you should have an external connection of some sort on every node, if nothing else for the purpose of alerts being generated to Isilon Support. There are other services that also will not work unless every node has an external interface. A management interface would suffice for this purpose.
That doesn't however mean that you have to have all of your clients connecting to all nodes. You can limit the node members within the SmartConnect pools. At that point the trade off is that you lose the available bandwidth from the additional six nodes. It depends on your workload as to whether or not that is an acceptable trade off.
In either case though, it is likely (depending on file size, anything over 128KB) that the file is striped across more than one node, so it will require communication over Infiniband to retrieve the rest of the file. So, the idea that you will have a 1/5 of 1/11 chance of the file being local, is slightly flawed. There is however a better chance of it being partially local by only connecting to the X nodes.
Isilon does have some performance troubleshooting commands that may assist in testing the environment.
Here is a knowledge base article that can help with that:
471726 : OneFS: Troubleshooting performance issues https://support.emc.com/kb/471726