# DeepIO

https://akougkas.io/work/deepio/

Author: Anthony Kougkas

Updated: 2026-11-12

## The line

DeepIO is a series of subcontracts with three national laboratories. [Argonne](https://www.anl.gov/) funded DeepMeta from 2022, then DeepIO and its 2026 renewal; [Pacific Northwest](https://www.pnnl.gov/) funded DFlow from 2022 to 2024; [Lawrence Livermore](https://www.llnl.gov/) funded WisIO from 2023 and IntentIO in 2026. Each one widened the line, and together they paid for students who wanted to work where I/O meets AI.

The question they share is older than any of them. Around 2020 every AI-for-science conversation ended at the same place: the GPUs were waiting for data, and nobody could say precisely why. Training, inference, and the diagnosis of both depend on a data path that crosses storage, memory, accelerators, and files, and DeepIO studies that path as one system.

## Where it started

[DLIO](https://akougkas.io/publications/devarajan2021dlio/) is [Hariharan Devarajan](https://grc.iit.edu/center/team/hariharan-devarajan/)’s benchmark. He built it with [Huihuo Zheng](https://www.alcf.anl.gov/about/people/huihuo-zheng) and [Venkatram Vishwanath](https://www.alcf.anl.gov/about/people/venkatram-vishwanath) at the [Argonne Leadership Computing Facility](https://www.alcf.anl.gov/), profiling cosmology, particle-physics, and climate workloads on Theta and turning them into a configurable emulator, and I was part of the [Illinois Tech](https://www.iit.edu/) team that co-advised him. The [CCGrid 2021 paper](https://akougkas.io/publications/devarajan2021dlio/) won the Best Paper Award. Hari’s [Stimulus](https://akougkas.io/publications/devarajan2022stimulus/) followed at CCGrid 2022, connecting scientific formats to AI frameworks and speeding up ingestion by up to 5.3x on Summit.

What DLIO became afterward is Argonne’s work, and mostly Hari’s. [MLPerf Storage](https://github.com/mlcommons/storage) from [MLCommons](https://mlcommons.org/) runs its workloads through `dlio_benchmark`, and the benchmark is maintained at [argonne-lcf/dlio_benchmark](https://github.com/argonne-lcf/dlio_benchmark), with [v2.0.1](https://github.com/argonne-lcf/dlio_benchmark/releases/tag/v2.0.1) as the current release.

## The students’ work

[TunIO](https://akougkas.io/publications/rajesh2024tunio/), [Neeraj Rajesh](https://grc.iit.edu/center/team/neeraj-rajesh/)’s IPDPS 2024 work, searched layered I/O configurations automatically. [Jie Ye](https://grc.iit.edu/center/team/jie-ye/) moved the question to inference. [Viper](https://akougkas.io/publications/ye2024viper/) at ICPP 2024 schedules and transfers model updates between concurrent training and inference through memory, cutting model-update latency by roughly 9x. [UnboxKV](https://akougkas.io/publications/ye2025unboxkv-io/) at IPDPS 2025, with [Bogdan Nicolae](https://www.bnicolae.net/) at Argonne, characterized how KV caches behave under concurrent transformer inference, and [PKAS](https://akougkas.io/publications/ye2026pkas/) at HPDC 2026 used predicted cache demand to schedule requests, up to 7.34x more throughput than the schedulers it was compared against.

## I/O characterization

The Livermore subcontracts fund I/O characterization, with [Kathryn Mohror](https://people.llnl.gov/people/mohror1) as the partner. [DFTracer](https://akougkas.io/publications/devarajan2024dftracer/), Hari’s SC 2024 tracer, captures AI workflow activity at 1 to 5% overhead. [WisIO](https://akougkas.io/publications/yildirim_wisio_2025/), [Izzet Yildirim](https://grc.iit.edu/center/team/izzet-yildirim/)’s ICS 2025 system, analyzes multi-terabyte traces through time, process, and file views and explains bottlenecks rather than only counting them: up to 805x more bottleneck coverage than a single view, and up to 11x faster analysis than the tools it was compared against. [WisIO v0.2.0](https://github.com/grc-iit/wisio) shipped publicly in July 2026. IntentIO, the 2026 follow-on with [Nikoli Dryden](https://ndryden.com/), extracts I/O intent from training runs with machine learning.

## Scientific AI as a data system

These threads meet at the same systems boundary. Training depends on the path from storage to accelerators. Concurrent inference depends on model movement, cache placement, and scheduling. Diagnosis depends on traces that connect application behavior to devices and files. DeepIO treats GPUs, memory, KV caches, storage, and provenance as one data path rather than a collection of isolated optimizations.

That systems view is DeepIO’s contribution to [IOWarp](https://akougkas.io/work/iowarp/) and [CLIO](https://akougkas.io/work/clio/). Context management for scientific AI has to carry data, model state, intent, topology, and history through the same runtime decisions.

## Where it stands

DeepIO is active. The 2026 Argonne renewal and the Livermore IntentIO award run this year, and [Isa Muradli](https://grc.iit.edu/center/team/isa-muradli/) and Jie are studying GPU-attached storage and dynamic tensor routing for multi-GPU inference. DLIO continues at Argonne with its community. The results belong to the people whose names are on the papers.
