⚠️🚧HANS — Hardware-Aware Neural Storage🚧⚠️

Caution

This repository is currently under active development and not yet production-ready!

HANS is an open-source, AI-native, cache-first storage system designed to maximize performance for modern AI workloads by being deeply aware of hardware — especially GPUs and accelerators.

HANS sits between AI compute and persistent storage, acting as an intelligent, distributed cache optimized for training, inference, and feature pipelines.

---

Why HANS?

Traditional file systems and storage layers are not designed for AI workloads:

• GPUs stall waiting for data
• Training throughput is limited by I/O
• Inference latency is unpredictable
• Edge devices fail due to VRAM pressure, not compute limits

HANS addresses these issues by making hardware state and AI access patterns first-class inputs to storage decisions.

---

Key Features

• Hardware-aware caching
  • GPU utilization, VRAM pressure, NUMA locality
• Memory-first architecture
  • RAM and NVMe prioritized over disk
• Edge-friendly design
  • Stable operation under tight VRAM and power constraints
• Asynchronous I/O
  • Built on modern Linux primitives (io_uring)
• AI-aware prefetching
  • Optimized for epochs, shards, and checkpoints
• Observability-driven
  • eBPF and hardware metrics inform decisions

---

What HANS Is (and Is Not)

HANS Is:

• A cache-first storage layer for AI workloads
• Optimized for GPUs and edge device constraints
• Designed for Ubuntu and Linux environments
• Open-source and extensible

HANS Is Not:

• A general-purpose filesystem
• A durable storage replacement
• A database or object store
• A universal drop-in solution for all workloads

Persistence is delegated to underlying storage systems.

---

Architecture Overview

graph TD
   subgraph AI ["**_AI Frameworks & Applications_**"]
       direction LR
       PT[PyTorch] --- TF[TensorFlow] --- TR[Triton] --- FS[Feature Stores]
   end

   subgraph Clients ["**_Client Interfaces_**"]
       direction LR
       POSIX[POSIX / FUSE] --- PY[Python API] --- NA[Native] --- S3[S3]
   end

   subgraph Core ["**Core Cache Engine (THIS DESIGN)**"]
       direction LR
       PL[[Placement]] --- PR[[Prefetch]] --- CH[[Chunking]] --- EV[[Eviction]] --- IO[[I/O]] ---  GPU[[GPU-aware]] --- FA[[Format-aware]]
   end

   subgraph Storage ["**_Storage Tiers & Backends_**"]
       direction LR
       RAM[(RAM)] --- NVMe[(NVMe)] ---  DISK[(SSD)] --- OBJ[(Object Store)] --- RFS[(Remote FS)]
   end

   AI --> Clients
   Clients --> Core
   Core --> Storage

Loading

Current Status

🚧 Early development / pre-alpha

HANS is currently under active development and not yet production-ready.

Initial focus:

• Core cache engine
• io_uring-based I/O
• GPU awareness
• Edge-friendly behavior

---

Roadmap (High-Level)

• Core cache engine MVP
• io_uring integration
• RAM + NVMe tiering
• GPU-aware placement
• Edge profile support
• Benchmarks vs ext4 / XFS
• Public alpha release

---

Getting Started

⚠️ Instructions will evolve as the project stabilizes.

Requirements

• Ubuntu 22.04+
• Linux kernel with io_uring support
• NVIDIA GPU (recommended)
• CUDA toolkit (optional but recommended)

Build (Early)

git clone https://github.com/ KAESTechnology/HANS.git
cd hans
# build instructions coming soon