Navigating the DoD IL5 "Valley of Death": Automating Hybrid Multi-Cloud Inference Server Deployment with the HPCMP High Security Platform (HSP)
Achieving secure, scalable AI/ML inference deployments within DISA Cloud Impact Level 5 (IL5) FedRAMP High compliance remains a significant barrier for many DoD research and mission-critical computing initiatives. The rigorous Authority to Operate (ATO) process, particularly at IL5, frequently creates a “Valley of Death,” hindering innovation through stringent security requirements, complex cloud configurations, and demanding compliance standards. Over the past three years, Parallel Works has successfully navigated this challenging landscape, obtaining DISA IL5 High Provisional Authorization (PA) for the ACTIVATE High Security Platform (HSP). As the first implementation partner under the DISA PA, HPCMP now provides the HPCMP HSP capability, enabling seamless execution and secure data transfer of export-controlled workloads across Defense Supercomputing Resource Centers (DSRC) and multiple IL5 cloud environments, including AWS, Google Cloud, and Azure. This presentation provides a detailed exploration of our journey through the DISA IL5 authorization process, highlighting critical technical and operational challenges. It covers user-owned multi-cloud Slurm cluster provisioning under strict boundary controls, challenges related to project timelines, access control complexities, and cost considerations faced by groups contemplating their own ATO processes. A central use case addressed is the integration and deployment of the GDIT PET team’s vLLM Triton inference server, enabling persistent, private Large Language Model (LLM) services integrated directly within an IL5-compliant chat interface provided by HSP. The session details solutions for token budgeting, observability, and efficient resource allocation, demonstrating how AI inference workloads are operationalized across complex hybrid infrastructures. Additionally, the presentation highlights our innovative approach to addressing the persistent challenges of Kerberos authentication and remote access within secure DoD computing environments. The ACTIVATE CLI facilitates secure, automated terminal access to IL5 resources without requiring Kerberos, significantly streamlining user authentication and accelerating operational workflows. Attendees will gain detailed insights into practical strategies, technical challenges overcome, and planned roadmap enhancements aimed at effectively merging rigorous compliance with agile, automated, scalable mission-critical AI/ML deployments.
IMPACT
Accomplishment: Achieved DISA IL5 FedRAMP High Provisional Authorization for the ACTIVATE High Security Platform, enabling seamless multi-cloud bursting for mission-critical AI workloads; Result: Reduced Authority to Operate (ATO) timelines by months to years—accelerating secure deployment of scalable AI/ML inference capabilities critical to DoD operational agility and mission readiness.
PRESENTER
Shaxted, Matthew
shaxted@parallelworks.com
847-254-0230Parallel Works
CO-AUTHOR(S)
Gary, Dr. Stefan
sfgary@parallelworks.comTorreira, Alvaro Vidal
alvaro@parallelworks.comMcQuade, Michael
mcquade@parallelworks.comNguyen, Quan
qnguyen@parallelworks.comLe, Louis
lle@parallelworks.comCATEGORY
AI/ML for HPC
SECONDARY CATEGORY
Mod, Sim & Analysis for Decision Making
SYSTEM(S) USED
Nautilus, Carpenter, IL5 AWS on HSP