On the plus side, there is a more recent version of Maui available in OPD as well... you can use this as an opportunity to upgrade Maui at the same time as you upgrade PBS to Torque.
I've cc'd oscar-users for reference.
Jeremy
At 09:21 AM 5/25/2004, Michael Edwards wrote:
After chatting with some folks at HPCS I am interested in removing PBS and putting in either Torque or the "grid" one that Bernard Li was packaging. You have mentioned a bug in the removal scripts a couple times and I was wondering what that was.
Thanks again for your help.
Original Message ----------------------- At 03:41 AM 5/25/2004, Carlos Vasco Ortiz wrote: >Hi Jeremy, >> >>You'll have to be more specific here... I'm not sure what you mean. >By resource manager extensions I mean the way to bring some information to >the scheduler not covered by PBS (the resource manager). In the >documentation about MAUI tells that this can be done via the flag -W in >the pbs scripts, but this has to be implemented in the configuration of >PBS. I don't know if this is already done in the OSCAR distribution of PBS. >> >> >>>2.- Since our cluster is quite heterogeneous in cpu speed (at least 3 >>>clock speeds) and we >>>have 2 different switches, we would like our parallel codes been >>>expanded along same speed cpus, all of them connected to >>>the same switch. I now how to do one of each contraint separatelly, by >>>means of the nodesets, but I don't know how to impose >>>both contraints at the same time. It is possible to define two nodesets, >>>and to select the nodes from the intersection of both nodesets? There >>>are other way to impose that? >> >> >>I think there is a way to do this w/ PBS. You will need to manually add >>"resource" descriptions on each node, and then specify all the resources >>on the job submission line (qsub) that you wish your nodes to match. So >>this basically means: >> >>Edit your /var/spool/pbs/server_priv/nodes file to add the resource >>descriptions and restart the server. >>Then, on your job submit command: >>qsub -l nodes=6:ppn=2:clockA:switchB job_script.pbs >>(assuming clockA and switchB are resource descriptions you've used) >> >About this point, I already have this solution implemented, but the user >has to choose the resource he wants to run, and once you have submited >your job, you can not use any different resorce. Sometimes you can need 2 >nodes, have selected the clock A, but the clock B is also OK and have some >nodes free before the clock A, so you are loosing some cicles...
If clockB is also ok, then there should be a third group representing all nodes that would be ok... such as clockAandB or something. PBS resources can certainly overlap each other.
> With nodesets from Maui, you can select any set, >but all the nodes you are given are from the same set. This is what I >need, but I have to use two nodesets at the same time, and I don't know if >this is possible.
I can't help you much here... I'm not too familiar w/ Maui nodesets.
>> >>Also, if installing a new OSCAR cluster, I recommend leaving out the >>included PBS package and using the Scalable PBS (a.k.a. Torque) >>instead. You can get it from OPD during installation, along w/ the >>CluMon package. >We have the cluster already installed and running in a production >enviroment. We are waiting a new server (with disk) for a new, different >development cluster, so we will be able to try Torque in this case.
If you end up trying to remove PBS on any existing cluster, please consult us here first. There is a minor bug in the uninstall scripts for it in OSCAR 3.0. Also, if CluMon is installed, it would need to be removed and re-installed as well. This is not a problem on a fresh installation with CluMon and Torque downloaded from OPD though.
Jeremy
>Thank you very much for your help, > >Carlos >-- >Carlos Vasco Ortiz (ITP Tecnolog�a y M�todos) >Tel: 34 91 207 91 21 [ITP-only internal ext.: 91 21] >Fax: 34 91 207 94 11 ><mailto:[EMAIL PROTECTED]>mailto:[EMAIL PROTECTED]
fix_pbs_uninstall.sh
Description: Binary data
