APPLICATION A5 / NATIONAL INFRASTRUCTURE
CLIO on national research systems.
A context runtime has to work where large-scale science happens. This story records where CLIO Core has been built and run, and how NeuroH5 uses it through its HDF5 connector.
Does the same context runtime build, deploy, and scale across the national machines scientists actually use?
National systems differ in CPU architecture, accelerators, interconnect, schedulers, and file systems. DeltaAI runs Arm-based NVIDIA Grace Hopper superchips; Aurora uses Intel GPUs and schedules with PBS rather than Slurm.
A runtime that only works on the development cluster is not infrastructure. Portability, repeatable deployment, and measurements under stated conditions are the bar.
The workflow, stage by stage.
- Ares · Illinois Tech
The development and performance-evaluation cluster where CLIO Core is built and run.
CLIO Core · Jarvis - DeltaAI · NCSA
NCSA’s AI system built on NVIDIA Grace Hopper superchips. The platform builds and runs there through an ACCESS allocation.
CLIO Core - Aurora · ALCF
Argonne’s exascale system with Intel GPUs. The Gnosis project record reports a CLIO Core port and scientific workload runs at the INCITE GPU hackathon. Public Jarvis recipes describe Aurora deployment configurations.
CLIO Core · SYCL
What the work has produced.
The Gnosis project record reports that the team ported CLIO Core to Aurora at the INCITE GPU hackathon and ran scientific workloads on the system.[1][2][3]
CONDITIONS · Institutional account accompanied by public deployment recipes; independently published scaling data is not linked here.CLIO Core builds and runs on the Ares cluster at Illinois Tech and on DeltaAI.[3][4]
NeuroH5 was onboarded through a transfer-engine-backed HDF5 VOL connector and loads without application code changes.[5]
The instruments involved.
Runtime, transfer engine, and HDF5 VOL connector deployed on each system.
Read more →02 / ComponentJarvis pipelinesPer-machine deployment recipes for Aurora, Delta, and other systems.
Open source ↗03 / ComponentCLIO Kit · Slurm, Lmod, JarvisServers an agent can use for scheduling and environment management on such systems.
Read more →Next steps.
- Publish Aurora scaling data with node configuration and workload stated.
- Publish DeltaAI transfer-engine measurements with GPU configuration, file system, and workload stated.
- Record the 2026 demonstration of one Jarvis pipeline deploying CLIO on DeltaAI and Aurora.
Who should get in touch
Facility staff, allocation holders, and application teams on ACCESS and DOE systems who want a context runtime measured on their machines.