# GSoC 2026: Out-of-Core Computing in Dagger - Initial Prototyping

**URL:** <https://discourse.julialang.org/t/gsoc-2026-out-of-core-computing-in-dagger-initial-prototyping/137161>\
**Category:** Julia at Scale\
**Tags:** distributed\
**Created:** [May 18, 2026, 11:16am UTC](https://discourse.julialang.org/t/gsoc-2026-out-of-core-computing-in-dagger-initial-prototyping/137161 "2026-05-18T11:16:30Z")\
**Posts on this page:** 1\
**Page:** 1

<div class="post-metadata">

**Author:** ![rainerrodrigues](https://sea2.discourse-cdn.com/julialang/user_avatar/discourse.julialang.org/rainerrodrigues/32/33229_2.png) [@rainerrodrigues](https://discourse.julialang.org/u/rainerrodrigues)\
**Post date:** [May 18, 2026, 11:16am UTC](https://discourse.julialang.org/t/gsoc-2026-out-of-core-computing-in-dagger-initial-prototyping/137161/1 "2026-05-18T11:16:30Z")

</div>

Hi everyone! I’m Rainer. I’m preparing a GSoC proposal for the **Out-of-Core Computing** project for Dagger.jl.

Coming from a background in mechatronics and embedded systems, dealing with strict hardware memory constraints and optimizing low-level routines (I’ve been working a bit with `CUDA.jl` recently) is a space I really enjoy, so this project caught my eye immediately.

Before I write up a massive architectural document, I want to get my hands dirty and prove out the core mechanics. My plan for the next few days is to:

1. Write a minimum reproducible example (MRE) that intentionally forces an Out-Of-Memory (OOM) crash in Dagger by mapping over massive arrays.
2. Draft a highly simplified local prototype to intercept the `Chunk` lifecycle and “spill/fetch” data to a local NVMe drive using serialization when a hardcoded memory limit is hit.

I was thinking of opening a GitHub issue to document the baseline OOM script and track my progress on the local-disk prototype.

Does this sound like a productive first step to the maintainers? Also, are there specific files in the scheduler’s source code you’d recommend I focus on first to understand how memory pressure is currently tracked per worker?

Thanks! Rainer