Thread: KUSU 102: Beowulf Cluster Patterns

2.1 Cluster Patterns

After working with so many customers over the last 10 years building Beowulf/HPC clusters... a few architectures keep recurring (please remember.. when I coined this terminology for clusters it was way back in 2004-2005).

I have shared these designs with customers over the years and many have adopted them and improved on them and have given me valuable feedback.

So in the spirit of the GOF in software development, I have created a few names for the most popular designs:

2.2 The Basic Cluster Pattern


This the most basic of all beowulf cluster setup. A head node, a complex of compute nodes and a GE switch. Very simple and straight forward. Recommended to use a simple non-managed switch to reduce networking configuration issues.


2.3 The Standard Cluster Pattern


In the Standard Cluster pattern, users add a high-performance compute network backbone to the cluster. This compute network typically supports MPI and nowadays I/O data like Lustre, GPFS or GFS. Common HPC fabrics used include Myrinet, Quadrics and Infiniband.


2.4 The Advanced Cluster Pattern


The Advanced pattern is typical of larger cluster setups or a small cluster with heavy IO requirements. A separate storage complex running a parallel filesystem like Lustre or GPFS is used to feed the compute nodes with data at high speed and low latency.


2.5 The Enterprise Cluster Pattern


In the Enterprise pattern a set of Login nodes are introduced. This is to offload large number of concurrent user logins into the head node. In this setup, the headnode then becomes a pure management node only. The Login complex will then handle all user login activities such as code editing, compilation and debugging and job submission.

Hi,

What do you think about the Clustered Storage with Distributed File Services ? I overhead that it doesn't need any storage head nodes. Isilon OneFS, Panasas PanOS, NetApp OnTAP GX, etc.

Thanks,

Clapton

Hi

Definitely the CLustered Storage with Distributed File Services are also a good option for the storage complex...

We have deploy Panasas before and it works very well and the customer actually bought another unit after several months... we also deployed it just by reading the manuals... this was several years ago.. of course we had much experience with file systems and could easily understand what the Panasas manuals wanted us to do...

But yes.. those are good options to consider...

What would be interesting would be to see how Red Hat positions itself in the HPC space... while Red Hat GFS is not the best HPC filesystem... but it is in there in RHEL, you get it for 'free', and if packaged and configured correctly... it could be a cheap low costs HPC filesystem for smallish clusters...