# LABIOS

https://akougkas.io/work/labios/

Author: Anthony Kougkas

Updated: 2026-08-27

## The question

Most I/O interfaces discard intent. They receive an address, a size, and an operation, then force the rest of the system to infer why the data matters and what should happen next.

LABIOS started with a different question: what if an I/O request carried its meaning with it? A label can bind an operation, a data reference, and routing metadata into something the system can inspect, schedule, and transform.

## Where it came from

This was the second half of my dissertation, “Accelerating I/O Using Data Labels.” [Hermes](https://akougkas.io/work/hermes/) had shown that middleware could manage the storage hierarchy; LABIOS asked what the middleware could do if requests told it what they were for. I wrote the [HPDC 2019 paper](https://akougkas.io/publications/kougkas2019labios/) with [Hariharan Devarajan](https://grc.iit.edu/center/team/hariharan-devarajan/), [Xian-He Sun](https://www.iit.edu/directory/people/xian-he-sun), and [Jay Lofstead](https://www.lofstead.org/) of [Sandia](https://www.sandia.gov/). It won the Karsten Schwan Best Paper Award, the conference’s top honor, out of 185 submissions.

The idea kept going. The provisional patent we filed during Hermes grew into [US Patent 11,630,834](https://patents.google.com/patent/US11630834B2/en), granted in April 2023 to Hariharan, Xian-He, and me. Four months later, [NSF](https://www.nsf.gov/) funded a follow-on OAC Core award to [Illinois Tech](https://www.iit.edu/) that brings in Jay and [Kathryn Mohror](https://people.llnl.gov/people/mohror1) at [Lawrence Livermore](https://www.llnl.gov/).

## The architectural bet

The label is deliberately small: an executable contract between an application and a data service. It introduces no new file format, and it makes no claim that metadata solves every storage problem.

That changes the shape of the runtime. Work can be decoupled from the application’s critical path, which is where the paper’s up-to-17x speedup over synchronous I/O came from. Different workers can handle storage, transformation, or movement. Policies can make decisions from explicit intent instead of reverse-engineering a stream of anonymous reads and writes. And because a label maps equally well to a POSIX file and a key-value object, one storage system could serve HPC and Big Data workloads at once.

![The LABIOS software stack: applications and agents issue labels through client APIs, a dispatcher schedules them, and Databot and pipeline workers execute them against storage adapters.](https://akougkas.io/assets/img/labios-stack.webp)

The LABIOS 2.0 stack. Clients create labels through C++, C, Python, or LD_PRELOAD interception; the dispatcher aggregates and schedules; Databot and pipeline workers execute against storage adapters.

## From result to software

The first system established the idea and its performance boundary in peer-reviewed work. The patent preserves the core representation. The current public prototype tests the abstraction against new workloads, clearer APIs, agent-facing tools, and the ordinary friction of open software.

The [`labios-2.0`](https://github.com/grc-iit/labios/tree/labios-2.0) branch identifies itself as 2.1.0-rc.1 for development and evaluation. NATS JetStream carries labels, DragonflyDB stages data, and Docker Compose on one host is the reference deployment. It is not a production release. Multi-node operation has not been demonstrated, and the agentic worker tier is planned.

## The people who built it

[Luke Logan](https://grc.iit.edu/center/team/luke-logan/) took the same separation of semantics from placement into memory with [MegaMmap](https://akougkas.io/publications/logan2024megammap/) at SC 2024. [Keith Bateman](https://grc.iit.edu/center/team/keith-bateman/) carried labels into schedulable data tasks in [DTIO](https://akougkas.io/work/dtio/). [Shazzadul Islam](https://grc.iit.edu/center/team/shazzadul-islam/), [Izzet Yildirim](https://grc.iit.edu/center/team/izzet-yildirim/), [Rajni Pawar](https://grc.iit.edu/center/team/rajni-pawar/), [Neeraj Rajesh](https://grc.iit.edu/center/team/neeraj-rajesh/), and [Hua Xu](https://grc.iit.edu/center/team/hua-xu/) are building the 2.0 prototype and the agent-facing storage work around it, including Acropolis and AgentStage, both accepted at eScience 2026.

## What changed in my thinking

I used to describe labels mainly as an I/O optimization. I now see them as a way to preserve intent across a system boundary. That expands the idea into pipelines, provenance, and context management, but only if the interface stays precise enough to test. Within [IOWarp](https://akougkas.io/work/iowarp/) and [CLIO](https://akougkas.io/work/clio/), the same principle connects an operation to the data, constraints, and transformations it carries across the runtime.

The next question is how few semantics a label needs before useful composition becomes possible, short of turning it into another opaque application protocol.
