Contents of this Page

The chip cluster consists of two main clusters: chip-cpu and chip-gpu, each with specific partitions and usage requirements. Each cluster is segmented into several partitions, each serving different access needs and user privileges. The partitions ensure equitable distribution and specialized access based on contribution and general availability.

chip-cpu Cluster

Open

Schematic of hardware partitions defined in chip-cpu.

Usage Requirements

All jobs on chip-cpu must specify the following for Slurm allocations

• Cluster: The Slurm cluster to request allocations from (chip-cpu) (--cluster=chip-cpu)

• Account: Your research group (--account)

• Time: Estimated job duration in minutes (--time)

• Memory: Estimated memory usage in MB (--mem)

Partitions

Partition Types

1. General Partition (general)

• Open to all users

• No preemption risk

• Includes former taki hardware (cpu2021 and high_mem nodes)

2. Faculty-Specific Partitions (pi_<umbcAccountName>)

• Reserved for specific faculty and their students

• Highest priority for partition owners

• Named pattern: pi_umbcAccountName

3. Match Partition (match)

• Contributors only

• No preemption between PI groups

• Resource limits prevent monopolization

4. Contributor Partition (contrib)

• Contributors only

• Includes all contributor and DoIT nodes

• Jobs may be preempted by partition owners with “grace” period of 600s (=10min)

• Resource limits prevent monopolization-chip CPUs,

5. Develop Partition (develop)

• Open to all users

• Intended for code development, testing, debugging, and short interactive jobs

• Maximum job time: 90 minutes

• Maximum resources per job:

  • 4 CPU cores

  • 50 GB memory

• Includes nodes:

  • c18-49

  • c18-50

Preemption

On chip-cpu, contributing groups are able to access the contrib partition (as discussed above). Every contributor also has a hardware partition named after the contributing group (e.g., pi_doit has the pi_doit partition). Any contributor may run on any machine within the contrib partition. A job running on hardware dedicated to a specific contributor (e.g., pi_doit) may be preempted (meaning that the job is cancelled in our case) by a member of that contributor’s group. Once a job is selected for preemption by Slurm, the job has 10 minutes to run before Slurm sends the signal to interrupt the job.

Usage Examples

Everyone can run

Contributors can run

chip-gpu Cluster

Open

Schematic of hardware partitions defined in chip-gpu.

Usage Requirements

All jobs on the chip-gpu cluster must specify:

• Cluster: The Slurm cluster to request allocations from (chip-gpu) (--cluster=chip-gpu)

• Time: Estimated job duration in minutes (--time)

• GRES: Generic Resources (GPUs) to be allocated (--gres=gpu:$numberOfGPUs)

• Memory: Estimated memory usage in MB (--mem)

Further, for those users attempting to charge PI-specific partitions, the following are also required:

• Account: Your research group (--account)

• Partition: Your research group (--partition)

Partitions

Partition Types

Default Partition (gpu)

• Open to all users

• If no partition is specified with --partition, this is the partition the allocation is charged to

• Includes all nodes

• Jobs can be preempted after 3 days

• Preemption “grace” period of 600 seconds (=10 minutes)

General Partition (safe from contributor preemption) (gpu-general)

• Open to all users

• Includes all nodes purchased with shared funds (NSF MRI/DoIT)

• Jobs can be preempted after 3 days

• Preemption “grace” period of 600 seconds (=10 minutes)

General Partition (gpu-contrib)

• Cannot be used for jobs by any users, only available for viewing the list of preemptible nodes

• Includes all faculty nodes on the GPU cluster (pi_X’s in the graphic)

Faculty-Specific Partitions(e.g, pi_doit)

• Contributors only

• Includes only the faculty-specific node(s)

• Jobs may be preempted by partition owners with GraceTime of 600s (=10min)

Preemption

On chip-gpu, every contributor also has a hardware partition named after the contributing group (e.g., pi_doit has the pi_doit partition). Any user may run jobs on these machines via the gpu partition, a job running on hardware dedicated to a specific contributor (e.g., pi_doit) may be preempted (meaning that the job is cancelled in our case) by a member of that contributor’s group. Once a job is selected for preemption by Slurm, the job has 10 minutes to run before Slurm sends the signal to interrupt the job. Users wishing to avoid this contributor preemption may charge the gpu-general partition (discussed above).

Usage Examples

Everyone Can Run

Contributor Can Run

For more information on how chip-gpu cluster contributors can access their specific GPU Hardware partitions, see this page: Accessing PI GPU Partitions