01 June 2012

169. ECCE, node hopping and cluster nodes without direct access to WAN

NOTE: I'm still working on this. But this sort of works for now. More details coming soon.
Update:  I've posted better solutions more recently for use with SGE. See aspects of e.g. http://verahill.blogspot.com.au/2012/06/ecce-in-virtual-machine-step-by-step.html and http://verahill.blogspot.com.au/2012/06/ecce-and-inaccessible-cluster-nodes.html

It's not easy choosing a good title for this post which describes the purpose of it succinctly yet clearly (sort of like this sentence), so here's what we're dealing with:

The Problem:
You can access the submit node from off-site. You can't access the subnodes directly from off-site. This post shows you how you can submit to each subnode directly. A better technical solution is obviously to use qsub on the main node. Having said that, with very little modification this method can also be adapted to the situation described here: http://verahill.blogspot.com.au/2012/05/port-redirection-with-eccenwchem.html

A low-tech, home-built example is the following:
-eth0 - Main  node - eth1 - (subnodes: node0, node1, node2, node3)/

eth0 is WAN, and eth1 is LAN. You can ssh to the main node from the 'internet'. There's no queue manager like SGE installed -- submission of jobs is done by logging onto each node and executing the commands there.

The example:
Main node:
    my.cluster.edu
Nodes:
   compute-0-0
   compute-0-1
   compute-0-2
   compute-0-3


The most obvious solution:
do port redirection. The downside is that it requires some technical skills of the users, and anything with networking and ssh is a PITA if they insist on using windows.


The smarter solution:
I was informed of this solution here.

ROCKS-specific stuff
If your cluster is running ROCKS 5.4.3 and you're having issues opening csh shells, just move /etc/csh.cshrc out of the way as a crude fix.
sudo mv /etc/csh.cshrc ~/
Don't forget to do this on all the nodes if they have local /etc folders! And it's not that easy -- the passwords aren't the same on the nodes as on the main node.
So, on the main node:
rocks set host sec_attr compute-0-1 attr=root_pw
sudo su
# cat /etc/hosts
#ssh 10.1.255.253
#mv /etc/csh.cshrc .
#exit

As the user you'll be running as, edit/create your ~/.cshrc

setenv GAUSS_SCRDIR /tmp
setenv ECCE_HOME /share/apps/ecce/apps
set PATH= (/share/apps/nwchem/nwchem-6.1/bin/LINUX64 $PATH)
setenv LD_LIBRARY_PATH /opt/openmpi/lib:/share/apps/openblas/lib
Repeat on all nodes.

SIDE NOTE: What I'm ultimately looking to achieve on the ROCKS cluster is front-node managed SGE submissions. Easy, you say? Well, ECCE submits a
setenv PATH /opt/gridengine/bin/lx26-amd64/:/usr/bin/:/bin:/usr/sbin:/sbin:/usr/X11R6/bin:/usr/bin/X11:${PATH}
and for some reason the ROCKS/CentOS (t)csh can't handle (most of the time) adding ${PATH} to itself since it's too long unless "s are added (it seems) i.e. this works consistently:
setenv PATH "/opt/gridengine/bin/lx26-amd64/:/usr/bin/:/bin:/usr/sbin:/sbin:/usr/X11R6/bin:/usr/bin/X11:${PATH}"
On my debian system either works fine (bsd-csh).

ANYWAY -- continue

Ecce:
Start setting it up using
ecce -admin



After that, do some editing in ecce-6.3/apps/siteconfig/

CONFIG.voldemort:
xappsPath: /usr/X11R6/bin
NWChem: /share/apps/nwchem/nwchem-6.1/bin/LINUX64/nwchem
Gaussian-03: /share/apps/gaussian/g09/g09
sourceFile:  /etc/csh.login
frontendMachine: my.cluster.edu

Machines:
compute-0-0 compute-0-0.local Dell HPCS3 Phase 1 Cluster AMD 2.6 GHz Opteron 40:5 ssh :NWChem:Gaussian-03MN:RD:SD:UN:PW




No comments:

Post a Comment