NCShare
The North Carolina Shared Research & Education collaborative infrastructure (NCShare) provides a shared computing environment that supports a common set of software and science drivers. NCShare provides researchers at participating institutions with access to computing tools and services to meet their research and instructional needs. NCShare's GPU-as-a-Service model provides access to 32 H200 NVIDIA GPUs in order to accelerate computational tasks that would be difficult to achieve through the use of local resources. All NC State-affiliated researchers are eligible for NCShare access, and select use-cases can leverage a high-priority allocation for increased resource limits.
NCShare complements NC State’s HPC system and can be a an excellent option for computational projects that require additional GPU capacity, including access to powerful H200 GPUs. NCShare may be especially useful for research workflows involving machine learning, artificial intelligence, large-scale data processing, modeling, simulation, or other GPU-intensive work.
To help researchers determine whether NCShare is an appropriate resource for their project, contact the RFS for consultation and guidance.
| Who is eligible? | NCShare is available to NC State researchers. |
| Is there a cost? | There is no cost associated with use of NCShare. |
| Contact? | For more information about NCShare, or to request a consultation, please contact the Research Facilitation Service. |
| Grant support | Researchers who need to describe NCShare resources in a contract or grant proposal may use the boilerplate text below.
|
Terms of Use
Please note: NCShare is an advanced resource with limited capacity for user support. NCShare is a shared and limited resource. Large or frequent GPU-inefficient jobs may cause users to be removed from the high-priority group.
Please keep high-priority jobs efficient or small/infrequent!
Because NC State operates under a monthly quota of computing operations on our allocation of NCShare's 32 H200s, access to this high-priority partition is provided using a gated model.
Access to baseline NCShare resources
New users are initially granted access to the normal NCShare partition. This includes access to compute nodes with full NVIDIA H200 GPUs, but comes with some restrictions on simultaneous jobs and job size.
Users of the normal partition are expected to adhere to efficiency minimums of 50% time-efficiency and 30% memory-efficiency (as assessed by NCShare) while running jobs in the normal partition. Users whose jobs do not adhere to this efficiency standard will be given one warning via email and their gpu access will be revoked thereafter.
Access to NCShare GPU resources
Following NCShare guidelines, after one month of efficient use of the system, users can apply for access to the high-priority partition, which provides them with enhanced resource limits and reduced wait times.
Researchers seeking access to the high-priority partition must demonstrate:
- Resource need: workflow requirements should clearly fit NCShare’s resources. Workflows better suited to NC State on-premise resources or those available though nationally funded resource providers (e.g. ACCESS) should not be run on NCShare’s high priority queue.
- Efficient use: GPU job efficiency be assessed according to the following metrics:
- Time-efficiency: portion of the job duration where GPU is actively used should exceed 70%
- Memory-efficiency: GPU memory (VRAM) efficiency should exceed 50%
- Time-weighted memory efficiency should generally exceed 50% when possible
Users of the H200 GPUs are allotted 1800 hours of compute time on the high-priority partition. When this compute time has been exhausted, users will be automatically removed from the high-priority group. Users may reapply for access, provided they have previously made efficient use of the resource.
Users may be removed from the high-priority group before this compute time has been exhausted if the jobs they have submitted fail to meet the following efficiency formula (evaluated on 30 day aggregate usage):
- 360 > (GPU-hours used) * (1 - GPU-time-weighted-memory-efficiency %)
- This formula will only be evaluated if aggregate use falls below efficiency expectation thresholds.
- Exceptions may be made on request for workflows that do not conform well to this formula
- If they are removed from the high-priority partition, users will still be able to use the NCShare system, but access will be limited to the normal partition
How to access NCShare
- Faculty and staff researchers may apply for access directly by completing this brief form
- Graduate students may not ask for access directly, but faculty and staff may request access for their graduate students via the form
- Set up a COmanage account
- Set up an ssh key OR go through Open OnDemand using your browser to access the system
- Optional: reach out to group-research_facilitation@ncsu.edu for access to the high-priority partition after 30 days of efficient use.