Arkansas High Performace Computing Center [hpcwiki]

Welcome as users on the AHPCC Pinnacle HPC system! As Pinnacle is a shared resource used by many concurrent researchers at any given time, it's very important to observe a few basic rules so as to ensure that everyone enjoys a favorable user experience…

ALL jobs are required to be submitted to compute nodes thru use of the SLURM Job Scheduler. Running computationally intensive tasks (or memory intensive tasks) on the login nodes is prohibited. Such jobs will be killed by administrators.

Job queues determine which specific resources a user can request, both in terms of run time and specific hardware. AHPCC requires that users select the appropriate queue for their job submissions at all times. Jobs that are incorrectly submitted to queues that don't match the application will be killed by administrators. Submitting to alternate queues just because the appropriate queue is busy is not allowed!

• Cloud Queue - If you are using a serial (non-parallel) task, submit to this queue as it is designated for this purpose specifically.
• Himem Queues - Don't submit jobs to this queue unless you truly need greater than 192GB of memory. If unsure, first test on a Comp queue to see if your task runs out of memory before migrating to the Himem queues.
• GPU Queues - Do not submit to the GPU queues unless you are certain that your application/task can effectively utilize the GPU hardware. AHPCC staff can help you evaluate your code if needed.
• Condo Queues - Users should only submit jobs to the actual condo queue if your research group has actually purchased resources for this queue! If you have condo resources, the AHPCC staff will inform you as to how to submit jobs to your specific resources. Non-owners can still utilize these resources for short periods of time by submitting to the Pcon06 queue. AHPCC has an agreement with Condo purchasers that non-owners can utilize these resources for up to 6 hours at a time, and with low priority to the queue.
• Trestles Queues - Kind of a “catch-all” queue…the nodes are almost 12 years old and only half as fast as the other offerings, but there are a LOT of them, and there are usually at least a few available at any time. We also offer a 288 hour walltime on these which may be of interest to some.

Research groups who have participated in the condo purchase plan and “own” resources that are integrated into the Pinnacle system should fully utilize those systems FIRST before assigning jobs to the general purpose queues. This helps to ensure that everyone has access to adequate resources to get their research done. AHPCC administrators continuously work to adjust queue parameters to best balance the needs of all researchers. Condo resources can be particularly beneficial of a group or user has a regular requirement to run tasks beyond the standard walltime maximum of 72 hours.

HPC Clusters like Pinnacle aren't ideally suited for interactive jobs (srun) where constant user attention is required. Please be mindful when running these interactive jobs that, when the resource is not being used, cancel the session so that others may utilize the resources until you need it again. Leaving these sessions running idle and unattended is strongly discouraged.

For inquiries about software installations, assistance with job scripts, assistance trouble shooting a failed run, etc…please provide any of the following information that would apply to the issue in an email to HPC-SUPPORT@listserv.uark.edu

• Location of job script on system
• A copy of the .out or .err file from SLURM
• Any additional bits of information that might help
• Specific link to software you wish to have installed

Users should be conscientious when submitting lots of jobs to the system. Remember, you are sharing the resource with up to 800 other users! While SLURM Job Scheduler parameters typically manage workflow for the system, it is considered bad etiquette to try and submit as many jobs (packing the system) as the resource will allow, or occupying all of a particular resource type (such as Himem or GPU nodes). Everyone's research is equally important in the context of system use. If you desire higher priority, please see one of the AHPCC staff about a possible equipment purchase thru our Condo program.