Panel: Building the NRP Ecosystem with the Regional Networks on their Campuses;
•Download as PPTX, PDF•
0 likes•105 views
This document summarizes a panel discussion on building the National Research Platform ecosystem with regional networks. The panelists discussed how their regional networks are connecting to and using the Nautilus nodes of the NRP. Examples included using NRP for deep learning and computer vision research at the University of Missouri, challenges of adoption in Nevada and potential solutions, and Georgia Tech's new involvement through the Southern Crossroads regional network. The regional networks see opportunities to expand NRP access and training to enable more researchers in their regions to take advantage of the platform.
Report
Share
More Related Content
Similar to Panel: Building the NRP Ecosystem with the Regional Networks on their Campuses;
Similar to Panel: Building the NRP Ecosystem with the Regional Networks on their Campuses; (20)
More from Larry Smarr
More from Larry Smarr (20)
Recently uploaded
Recently uploaded (20)
Panel: Building the NRP Ecosystem with the Regional Networks on their Campuses;
- 1. Slide 1 February 9, 2023 Slide 1 February 9, 2023 Building the NRP Ecosystem with the Regional Networks on their Campuses 4th National Research Platform Workshop February 9, 2023
- 2. Slide 2 February 9, 2023 Slide 2 February 9, 2023 www.thequilt.net
- 3. Slide 3 February 9, 2023 www.thequilt.net Regional Networks and Campus Partners with Nautilus Nodes Connected to NRP • CENIC • NYSERNet • OneNet • GPN • C-Light • MREN • OneNet • FRGP • PNWGP • Eastern Research Platform - Edge, OSHEAN, CEN, NYSERNet • Merit • Florida LambaRail • AREON • MOREnet • Albuquerque GigaPoP • KanREN • Network Nebraska • SD-REED
- 4. Slide 4 February 9, 2023 www.thequilt.net Panelists • Grant Scott, Director, Data Science and Analytics Masters Program, University of Missouri • Scotty Strachan, Principal Research Engineer, NevadaNet • Eric Buckhalt, Senior Director, Technology Strategy, Architecture, & Governance, Georgia Tech
- 5. Using NRP @ University of Missouri (Mizzou) Grant Scott Assistant Professor • Computer Science, Computer Engineering • College of Engineering Director, • Data Science and Analytics, MS Program • Institute for Data Science and Informatics Provided Data Science and Computer Science Learning Outreach using MU & NRP Jupyter • US Government Intelligence Agency • USDA ARS Long-Term Agroecosystem Research (LTAR) Data Managers • Tutorials at Regional and International conferences https://scottgs.mufaculty.umsystem.edu/ Evangelizing the use of Nautilus / NRP for research computing and STEM education
- 6. Using Nautilus For Research @ MU Grant Scott with mentees Alex Hurt, Anes Ouadou Case Study in Deep Learning for Computer Vision • Journal Publication Results in Weeks or days • Using State-of-the-Art Deep Learning Computer Vision Models for Satellite Imagery Data Sets Assisting numerous other researchers using ML & Deep Learning • Protein Structure Prediction • Robotics • UAS Navigation & Vision Tutorials & Workshops • Great Plains Network Annual Meeting 2022 • Getting Started on Nautilus and Kubernetes • SC 22 – GPN Booth • Scaling Deep Learning with Nautilus • Great Plains Network Annual Meeting 2023 • Getting Started on Nautilus and Kubernetes • Scaling Deep Learning with Nautilus
- 7. Scaling Deep Learning with Kubernetes on Nautilus • Using containerized model definition and list of jobs • Mounted persistent data storage to each pod • Each GPU job produces an associated trained model • Automation currently performed via environment variables and bash, but more sophisticated methods in development • Models are sync’d to Nautilus S3 bucket for later use in evaluation or other ML applications Dr. Alex Hurt Nautilus for Accelerated Research Computing
- 8. Deep Learning on Nautilus: By the Numbers Compute Intensive • Containerized deep neural architectures: 9 • Datasets trained on: 3 • PyTorch Models Trained: 27 • Training Epochs Completed: 8,100 • Iterations of Training Completed: 30,088,125 • Number of Images Processed: 240,705,000 • Trainable Parameters Optimized: 1,730,368,875 (1.73 Billion) Data Intensive • Data loading: 415.8GB • Neural Model Loading: 124,740 GB • Wall-Clock: ~77 days • Human Effort: <3 hours Dr. Alex Hurt
- 9. Deep Learning on Nautilus: By the Numbers Compute Intensive • Containerized deep neural architectures: 4 • Datasets trained on: 1 – Forest Fire Mapping • PyTorch Models Trained: 144 • Training Epochs Completed: 14,400 • Iterations of Training Completed: 515,550 • Number of Images Processed: 7,070,400 • Trainable Parameters Optimized: ~23 million Optimizing time to science discoveries • Wall-Clock: ~22 days • Human Effort: ~13 hours Anes Ouadou (PhD cand.)
- 10. Nautilus Supports Jupyter Hub & Jupyter Lab GEER Excels • Basic Python Programming • Used Nautilus Jupyter Lab • Trainees authenticated with GitHub accounts Basic Programming Course Parallel Programming for High- Performance Computing • Multi-Proc, Multi-Thread • MPI, CUDA Design and Analysis of Algorithms • C/C++ Programming • Analysis with Symbolic Python • Run-time empirical analysis with Python Computer Science Courses Parallel Data Analytics • Kubernetes Data Science Courses
- 11. CC* Team: Great Plains Regional CyberTeam • PI NSF Award OAC #1925681 • Helping the Great Plains region better leverage collective cyberinfrastructure resources • GPN Contributions to Nautilus and the NRP • GPN developing a CC* Regional Compute proposal, GP-ENGINE, to deploy NRP nodes into 8 locations across 6 states • Continuing our regional training and outreach at GPN and member REN conferences Proposed GP-Engine Sites
- 12. Contributions to NRP @ MU Grant Scott & Derek Anderson, non-NSF resources Four - 4x A100 Servers • 16x A100 GPUs • 4x 1TB System RAM • 4x 128 CPU Cores • 110,592 CUDA Cores • 1.2TB GPU RAM • Online Feb 3, 2023 gpn-fiona-mizzou-1 gpn-fiona-mizzou-2 gpn-fiona-mizzou-3 gpn-fiona-mizzou-4 Working with Materials Scientists • Physics constrained neural models, flat optics development • Automated characterization of nanoparticle energetic reactions Computer Vision for Geospatial & Remote Sensing Analytics • Wildfire Mapping from Space • Drought and Flood Mapping from Space • Illicit Deforestation Detection • Indigenous peoples mapping for healthcare delivery
- 13. Regional Perspective from NevadaNet 4NRP Workshop, La Jolla CA RENs & NRPs Ecosystem Panel February 2023 Scotty Strachan, Ph.D. Nevada System of Higher Education Principal Research Engineer Mountain Systems Scientist Co-PI, NSF-EPSCoR Track 1 + CC*’s @ScottySci sstrachan@nshe.nevada.edu
- 14. CLASSIC RESEARCH TECHNOLOGY USE & ADOPTION CURVE: - Why is this the case? [match of immediate need+capability] - How to make solutions widely useful? [culture shifts imminent] LAB GROUP PARETO CURVE RESOURCE UNITS
- 15. CAMPUS DISCOVERY current state of Nevada researchers? [national platform engagement not found]
- 16. Does it service a wide range of research workflows? Is it easy to get access? [onboarding flow] What is the learning curve to productivity? [grads] What is the support model? [facilitation] Is it a service commitment? [iteration rate/maint req] ECOSYSTEM COMPONENTS: VALUE PROPOSITION
- 17. Nevada example: one way to overcome friction is standardizing on federated access, (semi)automated onboarding/provisioning… …freeing up distributed research engineers for more high-touch project/data facilitation Coming now: a new Science/Teaching DMZ architecture more UNLV, UNR, DRI computing & data infrastructure goes here National computing & data infrastructure goes here
- 18. MANY OF OUR REGIONAL CHALLENGES ARE GEOGRAPHIC - I think a lot about the future of Wide-Area-Science [trust, community] - Roles of nrp components in the IoWT [data-to-action, workflows] - This will be driven by the regional networks [new mission areas] RESEARCH FUTURES IN THE WEST
- 19. Georgia Tech / Southern Crossroads (SoX) Relatively new to the NRP…. SoX is deploying nodes as part of CC* Area 2 regional awards that include several HBCUs in Alabama and Georgia FIONA nodes purchased as DTNs or perfsonar test points and are considering redeploying as nautilus nodes Goal is to enable research within our community and looking to assist faculty unfamiliar with the platform SoX is looking for toolkits, demos, training we can share, invite, host, etc. for our participants to help take advantage of the NRP SoX offers connectivity and space for those interested in hosting nodes
Editor's Notes
- Jen Leasure w/The Quilt – the national collaboration of regional research and education networks As I was preparing for this panel, I thinking the first national research platform workshop that took place about 5.5 years ago at Montana State University. Its exciting to be here today with you to share about how the PRP/NRP has scaled over this time as well as the collaborations that have been sparked due to this platform and to talk about future directions.
- Many of you are familiar with this map of our country’s regional research and education networks. A Number of Quilt member organizations are represented here at the meeting This map represents the backbone of each of these organizations What you don’t see are the last mile connections to approximately 1/3 of our country’s higher education institutions. Amy’s presentation this morning mentioned the Minds We Need vision – connecting every community college, TCU, HBCU, and minority serving institutions to our national R&E fabric which just one step toward the NSF goal for democratizing access to cyberinfrastructure resources. Our Quilt community feels strongly about this vision as we are positioned to serve as key partners in collaborating with lesser-resourced institutions and connecting them with the fabric of R&E networks as well as exposing and introducing them to the broader cyberinfrastructure ecosystem represented by all of you at this meeting today.
- Panelist represent regional networks in various stages of implementing nautilus nodes on campuses with their regional network partners