Are My node's working?

[sudeepk]

Yesterday night i have used the PelicanHPC live cd and connected  to 22nodes. When i ran the kernel_example command in octave it showed the same result as it showed for a single node. Does the nodes are sharing the work? how can i know that that the nodes are sharing the work or not?

[Michael Creel]

You can set up a ksysguard based monitor, as described on the forum (see the homepage, documentation section, which has a link). Or ssh into a node (e.g, ssh 10.11.12.2) and run htop to see the activity on the node.  Note that you need to edit the kernel_example script to specify that more compute nodes should be used, the default is to use only 2, if I remember correctly. While in octave, just type "edit kernel_example" and it should open up in the kate editor.

[sudeepk]

Thank you.. Till now i have worked on the cluster.These are the bench marks i have got.

node configuration
processors: intel dual core
ram: 1gb
network:standard lan.

3nodes:

gflops: 1.915
values in HPL.bat in /home/user/hpl2.X/bin/Pelican/
N=3000
Nb=64
p=1
q=3

10nodes:

gflops: 2.508
N=3000
Nb=64
p=2
q=5


i did not understand about the HPL.dat, so i have given some values. I also dont understand why the gflops are not increasing though i increased the nodes. And i think the network is very bad. The network is meant for normal connections in my college lab.

Another confusion what is 4 in

orterun --hostfile /home/user/tmp/bhost -np 4 xhpl

I have changed the values according to no of nodes...

[Michael Creel]

Hi,
Please check the homepage, in the documentation section. There is a link to a forum post that explains how to tune the HPL benchmark to get high values. The 4 after -np is the number is the number of MPI ranks to use. Please type "man orterun" to get more information on that.

HPL on PelicanHPC is included mainly as an example of making something run in parallel on the cluster, I personally don't ever  try to tune it to get high numbers, because my attention is elsewhere. I believe that the speed of the network is a limiting factor for most Beowulf style clusters that are what Pelican is usually used to make. Even so, you should be able to get much better numbers with 10 machines, if you are interested in doing so.

Cheers, M.

[sudeepk]

Thank you..

[calcpage]

I haven't had a chance to try the ksysguard info on the website.  What about ganglia or some other monitoring app for MPI?  I need to roll out my own version of pelicanHPC!  debian-live here I come....

Regards,
AJG

[sudeepk]

i havnt tried gangila. but i can recommend you Breakin which is a stress maker for the clusters which is better than the linpack hpc bench mark tester. The problem with pelican is not installable, and which is making me restricting me to use internet which is behind a proxy. May be i should work on it.