boilernet | Akshath Raghav R

BoilerNet Team 20! From L to R: Gokulkrishnan Harikrishnan, Akshath Raghav Ravikiran, Aneesh Reddy Poddutur, Gautum Kottayil Nambiar

Overview

This project aimed to develop a Network Attached Storage utilizing ESP32 MCUs. The aptly-named BoilerNet aims to enable In-Network-compute alongside swappable disks and compute workloads, neatly brought together by a cloud-hosted Dashboard interface. There are a few core ideas that form the foundation for this ”cluster” – Low Power Usage, Plug-and-Play Workloads, High Scalability.

BoilerNet won the “Senior Design Award” for Spring ‘25 through Purdue ECE.

Remark (Disclaimer)

Our reports/presentations are linked in the [/docs](https://github.com/AkshathRaghav/boilernet/tree/main/docs/). They contain in-depth views, reasonings and explanations regarding the design and its motivation. We also outline how we tested and ensured functionality. This repository is just meant to help anyone who wishes to 'refer' to our code/designs for their own use-case.

Top Level System Diagrams

Mechanical Design

The physical enclosure consists of six piece types, along with an optional divider: 1x Enclosure Lid, 1x Enclosure Base, 8x compute Enclosures, 8x compute Enclosure Lids, 6x 3mm diameter - 20mm screws, 4x 3mm diameter - 5 mm screws. You can find the related CAD files at /cad/ w/o dependencies. Reach out to Aneesh Poddutur for more information.

Physical Enclosure

Top Enclosure

Compute Enclosure

Assembled System!

PCB Design

This was our first time working on PCB Design, and had a few fly-wires hanging about. Below is one of the slides we usd in our Final Presentation to sound endearing ;)

You can find our PCB Designs at /gerbers/ and /kicad_projects! Reach out to Gautam Nambiar and Gokul Harikrishnan if you have any questions.

Network/Switch PCB

Compute Node PCB

Mishaps/Learnings!

Software

All the code for the MCUs are maintained within the /src folder. Each node’s build is maintained within it’s own ESP-IDF setup, and can be configured in isolation. /model_train contains the code for training the models and preparing weights to flashed through https://github.com/AkshathRaghav/boilernet/tree/main/compute_nodes. Please reach out to Akshath Raghav if you have any questions!

/src/fastapi_backend and /src/streamlit_frontend can be configured using the Makefiles. Try make venv; make setup; make run.
/src/network_node, /src/switch_nodes and /src/compute_nodes require idf.py configured and IDF_PATH set up. Remember, idf.py should be run in an ESP-IDF project directory, i.e., a directory containing a CMakeLists.txt file. idf.py fullclean; idf.py build; idf.py flash -port /dev/ttyUSB1 will be sufficient.
/model_train/extract.sh has the script for setting up ImageNet – only if you want to train your own subsets. Thanks to BIGBALLON for this script.
/model_train/train.py and /model_train/convert_and_header.py will set you up quickly with training your MobileNet or EfficientNet backbones with a head for the subset you’ve chosen, and then quantizing them (w/ additional calibration tuning). Results will be put into /model_train/models and /model_train/quantized_models. After you’re done training, put the chosen header (.cc) file into /src/compute_nodes/main and put this file name into Line 5 in /src/compute_nodes/main/CMakeLists.txt.