[Beowulf] Software installation across cluster

[Joe Landman]

On 05/26/2015 05:45 PM, Trevor Gale wrote:
> Hello all,
>
> I was wondering what solutions other use for easy software installation across clusters. Is there any method that is generally accepted to be the most effective for making sure that all the software is consistent across each node?

Many different methods, depending upon stateful/stateless modes.

For stateful modes (local installation, mutable boot drives), you have 
to worry about version/configuration drift, among many other issues.  
Here programmatic tools can help, but really, a simple shared file 
system (read only) across the stateful nodes has been my preferred 
mechanism when we use stateful systems.

For stateless systems, we normally have two types: hybrid and pure. 
Hybrid nodes will eventually mount state preserving directories, so 
curiously, they can make effective use of the shared mounting as 
indicated above.  Pure stateless mounts no external stateful resources, 
so the applications need to be local.  This is hard when the 
applications are sizeable.  FWIW, we keep our toolchain of about 800MB 
of app stacks, in /opt/scalable on our stateless nodes.  I've debated 
whether to use squashfs for this in the past, but its all one 8G ramdisk.

Stateful nodes are at risk if the server providing the file system goes 
offline.  Stateless nodes have limited space for apps.

For most things a remotely hosted /opt/apps or /apps etc. would 
suffice.  Use modules as indicated to handle environment bits.  As it 
turns out, the stateful single apps namespace makes installation of 
software across a cluster completely trivial.

With the use of docker/rkt and other similar things in HPC, expect the 
mechanisms to change, again.

-- 
Joseph Landman, Ph.D
Founder and CEO
Scalable Informatics, Inc.
e: landman at scalableinformatics.com
w: http://scalableinformatics.com
t: @scalableinfo
p: +1 734 786 8423 x121
c: +1 734 612 4615