IOWarp
IOWarp develops context-management infrastructure for AI-assisted scientific workflows. Its Context Layer for I/O (CLIO) architecture organizes data, metadata, telemetry, and computational state across storage tiers, then exposes that context to agents and scientific tools. CLIO Core supplies storage and runtime services. CLIO Kit connects agents to scientific data formats, schedulers, analysis software, and research datasets.
Research Problem
Scientific workflows span data formats, storage tiers, computing facilities, and institutions. Agents need more than file contents to plan multi-step work. They also need metadata, provenance, telemetry, and computational state. IOWarp treats those elements as context that can be ingested, placed, searched, and served through one architecture.
CLIO Architecture
CLIO has four public layers: CLIO Core for context storage and runtime services, CLIO Kit for scientific tooling, CLIO Agent for data-management operations, and CLIO Coder for research software development.
Conventional I/O
Applications interact directly with files and storage services
CLIO
A context layer coordinates data, metadata, and storage services
CLIO Core provides five groups of services:
- Context Assimilation Engine. OMNI-YAML descriptions bring binary, HDF5, Globus, and related sources into CLIO storage.
- Context Transfer Engine. Multi-tier buffering connects POSIX, HDF5, ADIOS2, MPI-IO, FUSE, and GPU-attached storage.
- Context Exploration Engine. C++ and Python APIs bundle, query, and retrieve context.
- CLIO Runtime. Dynamically loaded modules execute as coroutine-based distributed services.
- Context Transport Primitives. Shared-memory containers, allocators, synchronization, networking, and GPU-aware movement support the data path.
The design combines research lines from Hermes, LABIOS, ChronoLog, LabStor, and MegaMmap. Full architecture and API documentation is available at iowarp.ai/platform (opens in a new tab).
Agent-Facing Scientific Tools
CLIO Kit exposes scientific software through the Model Context Protocol. Public servers cover ADIOS, HDF5, Jarvis, Slurm, Darshan, ParaView, ArXiv, and other data and computing systems. The documented command is clio-kit mcp-server <name>, with configuration examples for Claude Code, Cursor, and VS Code.
The repositories and their release histories are maintained under github.com/iowarp (opens in a new tab). The full server catalog and installation guide are at iowarp.ai/clio-kit (opens in a new tab).
CLIO in DOE Genesis
Two DOE Genesis projects will extend CLIO capabilities through laboratory-led work. The Argonne National Laboratory project is led by Dr. Bogdan Nicolae. The Pacific Northwest National Laboratory project is led by Dr. Nathan Tallent. Across the two efforts, GRC will contribute CLIO Core, CLIO Kit, CLIO Agent, and CLIO Coder capabilities.
Research Results
AI-Driven Scientific Workflows
An eScience'25 poster (opens in a new tab) evaluated agent-driven workflows for EarthScope data, LAMMPS trajectories, ParaView visualization, and IOR deployment. It reports:
- EarthScope retrieval and visualization in about 2 minutes with Claude Code and Sonnet 4. The comparison agent did not complete the workflow without IOWarp context.
- LAMMPS BP5 analysis in about 2.5 minutes with Claude Sonnet 3.5, compared with about 12 minutes manually.
- IOR deployment in about 0.5 to 2 minutes across seven agent-model combinations and five prompt variations, compared with about 5 to 10 minutes manually.
Recent internal evaluations report 7.5 times faster end-to-end EarthScope analysis than the manual workflow, successful configuration in all four tested agent-model combinations, and a fastest run of 24.8 seconds with split plan-execution on local models.
Memory and Storage Convergence
MegaMmap manages DRAM and storage as tiered software distributed shared memory. Workload-aware eviction and prefetching reduced peak DRAM use by as much as 2.6 times at competitive performance in the SC'24 paper (opens in a new tab). Its tiered-memory design informs CLIO's Context Transfer Engine.
Semantic Search and Data Staging
Acropolis adds adaptive-depth content indexing to a tiered filesystem and exposes it through an MCP server. AgentStage uses an agent's reasoning phase as a window for moving requested data into place. Both papers appear in the eScience 2026 program (opens in a new tab).
Reproducible Deployment and GPU Data Movement
Jarvis describes portable scientific software stacks through a resource graph and automates deployment, monitoring, and teardown. Read the PDSW'24 paper. A separate eScience'25 poster characterizes GPU-accelerated data transfer through NVIDIA NIXL for AI workloads.
Browse all IOWarp publications.
Deployments and Scientific Workflows
CLIO Core has been built on Ares at Illinois Tech, NSF DeltaAI at NCSA, and ALCF Polaris and Aurora. At the INCITE '26 GPU hackathon, the team ported CLIO Core to Aurora, demonstrated scalability to 512 nodes, and ran scientific workloads at up to 128 nodes. A Context Transfer Engine-backed HDF5 VOL connector supported NeuroH5 and MiV-Simulator workflows. These deployment results come from IOWarp's Year 2 report to NSF.
Published poster demonstrations cover EarthScope geoscience data, LAMMPS molecular dynamics, and ParaView fluid visualization.
LAMMPS trajectory analysis through ADIOS tools
Pipe-flow visualization through ParaView tools
CLIO GenomIO
CLIO GenomIO applies AI agents and DNA language models to the gaps that repetitive sequences and low coverage leave in draft genome assemblies. The prototype accepts a gap and its flanking contigs, retrieves relevant genomic context, and returns a predicted sequence with a confidence score. LangChain-based planning agents orchestrate DNABERT-2, GROVER, Gena-LM, and Nucleotide Transformer over a retrieval layer built from a genomic corpus.
The work was developed with the High Throughput Sequencing Core at the Institute for Food Safety and Health (IFSH) at Illinois Tech, which provided sequencing facilities, bacterial genomes for pilot studies, and domain expertise. It is paused at the prototype stage, and the implementation is available on GitHub (opens in a new tab).