The Center for Research Computing continues to expand our Generative AI offerings from existing GPU infrastructure and HPC support now to an experimental shared LLM service platform. This platform, which is a deployment of the popular Open WebUI project, utilizes the CRC’s access to NVIDIA’s novel GH200 Grace Hopper Superchip to provide a chat interface similar to other generative AI engines such as OpenAI ChatGPT or Anthropic Claude. This beta resource is now available for all CRC users at this link: https://openwebui.crc.nd.edu (on campus or VPN, Notre Dame Google login required).

When using the platform, users will have the ability to test different open weighted multimodels directly on CRC hardware rather than requiring a commercial model subscription such as ChatGPT. This testbed is preloaded with the popular Llama3, Llama Vision, and Codellama models from Meta. All prompts, uploaded documents, and generated responses will stay on the CRC hardware, and will not be sent to any commercial provider or used to retrain commercial models. 

We welcome our users to experiment with these available models or to request other open weighted Ollama compatible models to be added to the system! Platform availability is offered as a beta test to evaluate broader campus or CRC use cases, so there may be outages and/or breaking changes as we continue to refine the system and monitor usage levels. If you have any questions about Open WebUI features available on our instance, or if you want to request additional models be added, please email crcsupport@nd.edu.

Efficient file system input/output (I/O) management is essential for running workflows smoothly, minimizing delays, and ensuring optimal use of system resources. Poor I/O practices can lead to bottlenecks, increased load on shared systems, and wasted time. By following these best practices, you can improve performance and avoid common pitfalls in data-intensive tasks.

1) Write to Local Disks First

Optimize your workflows by writing temporary data to local server disks whenever possible. Utilize the $TMPDIR directory provided by the Grid Engine job scheduler in your job scripts. This practice reduces the strain on shared file systems, enhances performance, and accelerates job completion. Once your workflow is complete, transfer only the essential results back to shared storage. Adopting this approach minimizes shared storage load, prevents unnecessary delays, and ensures efficient workflow execution.

Example for batch jobs:

cd $TMPDIR 
cp /path/to/input_file . 
my_program input_file > output_file 
cp output_file /path/to/shared-filesystem

2) Minimize Small Writes

Frequent, small writes can slow down workflows and strain storage systems. Each write operation has inherent overhead, especially on shared file systems. Instead of writing small amounts of data frequently, buffer information and write it in larger chunks. This technique is particularly important for tasks that generate logs, intermediate results, or outputs requiring frequent updates. Writing in bulk typically improves performance.

3) Optimize Output Resolutions

Tasks that generate large datasets can often benefit from reduced detail for intermediate results. Save high-resolution data only for the final outputs or critical stages of your work. Intermediate stages can use coarser resolutions to save space and processing time. This strategy is especially valuable in simulations, data analysis, and imaging workflows, where intermediate results may not need the same level of precision as the final output. By limiting detail when it isn’t required, you can save both time and storage resources.

This strategy is especially valuable in simulations, data analysis, and imaging workflows, where intermediate results may not need the same level of precision as the final output. By limiting detail when it isn’t required, you can save both time and storage resources.

4) Clean Up Conda Environments

Software environments and their associated files can gradually consume large amounts of disk space, leading to clutter and reduced efficiency. To free up storage and maintain an organized system, it’s important to regularly clean up unused packages, environments, and caches.

For users of Conda, the following command can help:

conda clean -a

By efficiently managing your environments, you can ensure smoother workflow execution and avoid delays caused by unnecessary or outdated files. This practice is particularly beneficial for those who frequently use package managers like Conda, as they can quickly accumulate redundant data over time.

5) Clean Up Temporary Data

Temporary files generated during a task can consume a significant amount of disk space if not managed carefully. Ensure your workflows include cleanup steps to remove temporary data when it’s no longer needed.

By cleaning up regularly, you can avoid unintentional storage overuse and ensure a tidy working environment. Automating this step in your workflows ensures consistency and prevents errors caused by leftover files.

6) Seek Assistance

If you experience challenges with storage or I/O efficiency, feel free to reach out to CRCSupport@nd.edu. Our team can provide personalized recommendations, pinpoint inefficiencies, and assist in optimizing your workflows.

To help us better understand your needs, please include specific details about your tasks, such as the types of files you’re accessing, the frequency of read/write operations, and any performance issues you’ve encountered. Providing this information enables us to offer targeted solutions and help you achieve the best possible results.

Conclusion

Adopting these I/O best practices can significantly enhance the efficiency of your workflows. By optimizing how data is read, written, and managed, you can minimize delays, reduce resource contention, and ensure smooth execution of your tasks. Good I/O practices not only improve your performance but also contribute to the sustainability of shared systems and resources. 

Every Wednesday and Thursday
2:00 – 3:00 p.m.
812 Flanner Hall (map)

The CRC offers multiple training opportunities for both new and existing users. We periodically provide short courses and other learning opportunities, which are advertised on our website and through email lists. In-person office hours are held every Wednesday and Thursday from 2:00-3:30 p.m. in Flanner Hall, room 812, on a first-come, first-served basis. You can also arrange a Zoom meeting at your convenience by emailing CRCsupport@nd.edu with your availability. We recommend bringing a laptop to in-person sessions.

For more information, please contact CRCsupport@nd.edu.

Aerospace & Mechanical Engineering
716,052 CPU hours

Civil & Environmental Engineering & Earth Sciences
506,410 CPU hours

Civil & Environmental Engineering & Earth Sciences
423,579 CPU hours

Aerospace & Mechanical Engineering
301,956 CPU hours

Chemical & Biomolecular Engineering
269,287 CPU hours

Chemical & Biomolecular Engineering
201,582 CPU hours

Chemical & Biomolecular Engineering
180,056 CPU hours

Chemical & Biomolecular Engineering
163,999 CPU hours

Chemical & Biomolecular Engineering
162,007 CPU hours

Physics
158,461 CPU hours