20 czerwca 2011 09:29
Sorry for the noob questions. We're likely to be setting up our first HPC platform soon. I'm not certain what platform this will be based on yet, but I'm obviously looking at Windows HPC edition. The cluster will be primarily for engineering simulations, mainly mathematical. Most of the software we use runs on Windows, so would work on the cluster.
As I've never done anything like this before, I have a couple of questions about how this would work.
Does the software have to be specically written to work in an HPC cluster, or can you take off the shelf software, install it on the cluster, and it will just work?
Will the software being used for the simulations have to be installed on each node in the cluster, or just the head node?
How does an end user submit jobs to the cluster? Is it via a client, web page or other means?
When a job is submitted, is it truly spread amongst all nodes in the cluster, or just submitted to one node for processing? Is this also dependant on the software itself?
Do you need an HPC license for each node in the cluster, or just the head node? Are the license types different for a head node and a compute node?
As I'm just starting to look into this, I would really appreciate some feedback.
20 czerwca 2011 10:20
Have you read the FAQ at http://www.microsoft.com/hpc/en/us/product/high-performance-computing-faqs.aspx?
There should be plenty of information already available to you but please don't be afraid to ask if you still can't find what you are looking for.
20 czerwca 2011 10:48
Excellent Q&A. Thanks Brian. I would think even the HPC Express version would get us going at the moment. We only have the capacity to scale to 12 compute nodes, each with 2 x 6C X5650 Xeons and 32GB RAM, but will start with 4 nodes.
This does answer a lot of my questions, but it doesn't go into too much detail about getting software working across the cluster. It just mentions developers using the SDK to develop apps that work. Does this mean standard Windows apps won't work, or just won't distribute the workload across the compute nodes? Are MS relying on vendors writing versions of their apps that are HPC aware?
29 czerwca 2011 11:43Hi tman24!
It's not an "official" answer but maybe it could help you:
" It just mentions developers using the SDK to develop apps that work. Does this mean standard Windows apps won't work, or just won't distribute the workload across the compute nodes?"
Normal Windows apps is working on HPC but you should take care of distributing the workload (e.g. parametic sweep applications, use windows shares for data access). You could also use MPI based application and the dyrad framework to do it.Also possible to send SOA jobs on HPC (this cases you write WCF application which will depend on the HPC SDK). In that cases the workload will spread on your cluster without changing anything.
If you need more details the following links could help you: