As discussed at the group's monday-meeting, we should gather information how we think the different SNIC/SNAC resources work. I'll start, but then I'd like the others in the group to add their experiences, thoughts and impressions.
SSH. No problems unless one wants to use AFS/Kerberos, which might be trickier.
PBS. No particular problems. One can check 'showstart' and also project info (though the last one is often outdated).
Access with AFS, if one whish to use that. Also, HPC2N uses KFS, a paralell filesystem. To make things easier, one can symlink the normal and the paralell systems.
OK. Some stuff like Octave and Xmgrace are missing.
Support is OK. Info by email/webpage is rather sparse.
SSH. No problems.
PBS. No problems.
Nothing special. No quotas. Scratch disk available.
Very good. Most stuff exists and if one misses something, it's just an email away. The support install almost whatever you need.
Good. Most info you need is on the webpage and the admins mail the users whenever something is happening.
Very good. Quick replies.
SSH. No problems.
SGE. One can only have 12 jobs in the que, which is really bad!
Uses quotas, which is set to 4 GB. No warnings will be sent to warn you of a full quota.
Missing Octave and Xmgrace.
OK, missing info on Math libraries.
Have become better, but before this was problematic.
PDC uses Kerberos for authentication and AFS for uploading files. This would have been ok if the other supercomputing centers used the same, but now it's a bit of a hassle. Loggin in requires several steps. If you don't set the Kerberos tickets long enough, your jobs might crash!
PDC uses the Easy scheduling system. Works ok, but it's tricky to find out when a job's going to start. Also, why not use PBS like all the others?
AFS. Also, quota is set to 500MB by default. This can be increased by asking for more.
Good. The support has put down a lot of effort to get Kerberos/AFS working for me.
Swegrid is quite different from the other resources in that it is a grid system. Simply speaking it consists of three parts: client software, resource manager and compute clusters. Jobs are submitted with the client software and sent to the resource manager that will find a suitable cluster for your job. Once this is done the job will be sent to a suitable compute cluster where the calculation is performed. After your job is finished you must retrieve the results. If you don't do this fast enough, and you're not using the local storage systems, your job might be deleted!
At the moment I use Green for running the client software. This works OK, but it's not the idea of grid computing that one has to have a cluster in order to submit jobs to other clusters. A client for Linux exists, but not for OS X.
Different clusters may use different queing systems locally. Most seems to use PBS.
Since one doesn't know what system your job will be run on, the safest way to go with your own codes is to compile it statically on a system that has the same architecture. When submitting your jobs you can specify what architecture you want to use. At the moment, Swegrid only consists of the X86 architecture.
Depends on the cluster. On some it's good, on others not so good.
Ok. There is a manual for Swegrid, which is ok but needs updating. There is also a manual for Nordugrid and the client software, but this is heavy reading. The Swegrid webpage doesn't work very well lately. Fortunately one can check the Nordugrid webpage instead.
Good. Support tickets are divided among the different Swegrid parts. See the other resources…