Table of contents
High-performance computing (HPC) is the use of parallel processing for running advanced application programs efficiently, reliably and quickly.
HPC Quick Start.pdf
This cluster was funded by what is commonly called the "Condo Compute Model".
The Premise core infrastructure is provided for UNH researchers. The Premise core infrastructure includes: racks, power distribution, cooling, InfiniBand Network mesh, Lustre file storage, a head node and four compute nodes. You may thank; UNH RCC, UNH Central IT, and the Research Office for providing this infrastructure.
Users compete equally for available resources in a "shared" job queue. All Premise users are expected to play nicely. An HPC Advisory Board has been created to provide RCC with direction on enforcement for the common good. Every attempt will be made to utilize available resources fairly.
Your budget should include "HPC buy-in" funding to satisfy your projects minimum needs. There is no other way to garantee the required resources will be available when you need them. The "Hardware Description" below defines three standard node configurations (approximate price of R730 nodes on 12/1/16): base ($8k), hi-ram ($12.5k), and gpu ($13.5k). In 2018 Dell discontinued the R730 model, which was replaced by the R740 at significantly higher costs. Contact RCC for current pricing for your proposal budget. Your grant retains ownership of any hardware you purchase.
Owners are provided a restricted job queue with priority scheduling on any hardware they own. When no owner priority work exists "shared" queue jobs may be scheduled on the idle hardware. Owners should expect active jobs to complete in a "reasonable amount of time". This might cause wait times for some priority jobs.
The Premise cluster is an HPC made up of:
See purchase history for more detail.
Premise has 31 compute nodes with two CPUs per node for a total of 28 CPUs (not including 3 nodes with 10 additional CPUs). Each 12-core "Intel(R) Xeon(R) CPU E5-2680 v3 @2.50GHz" is rated at 356.50 double precision GFlops.
(Total CPU performance) = (31 node) * (2 CPU/node) * (356.50 GFlops/CPU ) = (22103 GFlops) = (21.58 TFlops)
Ten nodes of the Premise cluster contain NVidia K80 GPU cards. NVidia specs claim upto 2.91 Teraflops double precision for a single K80 card.
(Total GPU performance) = (10 GPU) * (2.91 TFlops/GPU) = (29.1 TFlops)
(Combined Performance) = (Total CPU performance) + (Total GPU performance) = (21.58 TFlops) + (29.1 TFlops) = (50.68 TFlops)
Premise is managed by UNH Research Computing Center staff. Please email administrative, technical requests to: firstname.lastname@example.org
The focus of the Premise cluster is to support UNH Research. If you are seeking academic student HPC experience we currently suggest using XSEDE resources. For more information on XSEDE please contact the UNH Campus Champion "Scott Valcourt" Scott.Valcourt@unh.edu
For researchers within COLSA, additional information and instructions can be found here.
Create a Premise account by emailing UNH Research Computing Center staff at: email@example.com
For account creation requests we suggest the following information:
Full Name: Email and phone: Requested login id: Research Group: Expected use case / Research area: (Other relevant information)
Examples of other relevant information:
The only way to connect to Premise is by using a SecureSHell (SSH) encrypted connection. You will need an SSH Client program on a your internet connected computer to reach Premise. Often this can be done on your local command line by typing:
You may need to install an SSH Client on your computer.
HPC software is most often field specific. You probably have a better idea of where to look for relevant software tools in your field that we do, but you are welcome to ask RCC what we might know.
If you are bringing your own source code or use common Linux tools they may already exist on the cluster. Some software packages may be available as "modules" (more information available below).
You should not run your programs directly on the Premise login head node.
First copy any required data onto the Premise system, preferably in a subdirectory of your home area. Your same home area is mounted on all of the Premise nodes.
Once you have the application and necessary data ready you will submit it as a job into the batch system using Slurm. For details on using Slurm start with the local slurm usage notes.
You may ask other Premise questions via email to firstname.lastname@example.org. This is RCC's general support email, so please indicate that your question is related to "Premise".
RCC chose XDMoD to visualize Premise usage. The Premise XDMoD webpage may be viewed when on campus or using the VPN.
The batch system is used to schedule jobs to run on the compute nodes. Each command supports a --help option and has a Unix man page.
More information can be found here:
Common comands include:
Used to select specific software packages and exact versions. See the official module man page for more information.
Common commands include:
Users of Matlab on the Premise compute cluster should not run graphically on the Premise head node. Unlike running on your desktop, matlab jobs must be submitted to the Slurm job queue. A helper script has been created to submit your matlab.m scripts for you.
Run your matlab script with:
[rea@premise ~]$ module load matlab [rea@premise ~]$ sMATLAB.py matlabscriptname.m
Use "sMATLAB.py --help" to describe available options and defaults.
Adding the "--verbose" option to sMATLAB.py displays both the Slurm sbatch command line and helper job script that is being generated for you. This could be used as a starting point for users wishing to create their own Slurm scripts.
Note that Matlab does not automatically use all the cores on a node, or split a job accross multiple nodes for you. These features must be coded into your scripts. You may find the following links helpful:
Here is an example matlab script utilizing "parfor" to iterate work accross all the cores on one node. Using Matlab on more than one node is not supported. Premise nodes currently all have 24 cores.
1 2 3 4 5 6 7 8 9 10
parpool(str2num(getenv('SLURM_JOB_CPUS_PER_NODE'))); % workers=24 cores per Premise node tic % start timer ticBytes(gcp); % time should include distribution transfers n = 1024; A = zeros(n); parfor (i = 1:n) % Distribute these "n" iterations over workers in parpool. A(i,:) = (1:n) .* sin(i*2*pi/1024); end tocBytes(gcp) % timer should include collection transfers toc % stop & display elapsed time.
For installed COLSA related software, please see here.