This is an old revision of the document!
schumann.coastal.udel.edu
Vendor: TeamHPC #79116
UD Property Tag: 143989
| Operating System | Architecture | RAM | |
|---|---|---|---|
| head node | CentOS release 4.3 (Final) | 1 x AMD Opteron 275 (2 cores) @ 2200 MHz | 2000 MB |
| storage subsystem node | EonStor Infortrend Appliance | 1 x IBM PowerPC 750FX @ 800 MHz | 512 MB |
| (15) compute nodes | CentOS release 4.3 (Final) | 2 x AMD Opteron 275 (2 cores) @ 2200 MHz | 4000 MB |
Inventory:
|
|
Filesystems
The schumann.coastal cluster has a massive 10 TB fibre channel RAID5 disk array attached to it. The array is partitioned into two filesystems that are then mounted on the head node of the cluster:
/home: A 2.5 TB volume used for user home directories./monolith: A 7.5 TB volume for a large data store required by the owner
The /monolith filesystem is only visible on the head node; none of the compute nodes have access to it. However, the /home filesystem is NFS-shared to the compute nodes across the Ammasso RDMA Ethernet interconnect. This network fabric runs at gigabit speed, with enhanced, off-processor data streaming to the adapters.
This may seem odd at first glance: typically one would want to reserve the RDMA network for MPI-heavy streams. At the moment this cluster will probably see heavy use of Gaussian, constrained to single nodes, so the primary bottleneck will be the writing/reading of checkpoint files in the users' home directories. Sharing /home over the Ammasso interfaces makes sense in this situation.
Gaussian
The cluster has Gaussian '03 revision C02 installed on it. Gaussian jobs should be submitted to the compute nodes by means of gqueue, which (behind the scenes) creates GridEngine queue scripts and submits them for you, handling all of the sticky details:
% ls
bin H2O.com
% gqueue --nproc=4 H2O.com
Your job 5 ("H2O.gqs_XX2eKh6z") has been submitted.
% ls
bin H2O.chk H2O.com H2O.log
When selecting the number of processors you'd like for a job, keep in mind that as things are currently configured, Gaussian will only run on a single node, implying anywhere from 1 to 4 processors per job.