DeepSeek Open-Sources Fire-Flyer File System (3FS): Revolutionizing AI Workloads with High-Performance Distributed Storage

DeepSeek’s latest open-source innovation, ​Fire-Flyer File System (3FS), redefines distributed storage for AI workloads. Designed to address the escalating demands of modern AI training and inference, 3FS combines cutting-edge hardware optimization with groundbreaking architectural design. Released today as part of DeepSeek’s Open Source Week, this system is engineered to deliver unparalleled performance, scalability, and ease of use for developers and enterprises.

​Core Features of 3FS​

​1. Decoupled Architecture for Unmatched Performance​

3FS employs a ​disaggregated architecture​ that pools thousands of SSDs and hundreds of storage nodes, enabling applications to access storage resources without location constraints. This design:

  • ​Maximizes throughput: Aggregates SSD bandwidth and RDMA network capacity for seamless data access.
  • ​Ensures scalability: Supports clusters with 180+ nodes, achieving ​6.6 TiB/s read throughput​ under stress testing.

​2. Strong Consistency with CRAQ Protocol​

Leveraging ​Chain Replication with Apportioned Queries (CRAQ), 3FS guarantees ​strong data consistency​ while simplifying application logic. Developers can focus on building AI models without worrying about synchronization complexities.

​3. Universal File Interface​

3FS provides a ​familiar file interface​ backed by a stateless metadata service powered by transactional key-value stores like FoundationDB. This eliminates the need to learn new storage APIs, accelerating adoption and integration.

​4. AI Workload Optimization​

Tailored for AI workflows, 3FS supports:

  • ​Data Preparation: Efficiently organizes hierarchical directories and manages intermediate outputs.
  • ​Training Acceleration: Enables random access across compute nodes, removing the need for data prefetching or shuffling35.
  • ​Checkpointing: High-throughput parallel checkpointing for large-scale training tasks.
  • ​Inference Optimization: Cost-effective ​KVCache​ solutions with ​40 GiB/s read throughput, outperforming traditional DRAM caches.

​Performance Benchmarks​

​1. Peak Throughput: 6.6 TiB/s​

In a 180-node cluster (each with 16×14TiB NVMe SSDs and 2×200Gbps InfiniBand NICs), 3FS achieved ​6.6 TiB/s aggregated read throughput​ under concurrent client requests, even with background training traffic.

​2. GraySort Benchmark: 3.66 TiB/min​

Using the Smallpond framework on 3FS, DeepSeek sorted ​110.5 TiB of data​ in ​30 minutes 14 seconds​ (25 storage nodes + 50 compute nodes), showcasing its efficiency in large-scale data processing.

​3. KVCache Efficiency​

For LLM inference, 3FS delivers ​40 GiB/s peak read throughput​ per client node while maintaining stable garbage collection IOPS, making it ideal for real-time AI applications.

​Integration with Smallpond Framework​

3FS synergizes with ​Smallpond, a lightweight data processing framework built on DuckDB, to streamline operations like SQL queries and data transformations. Together, they enable:

  • ​PB-scale data handling: Seamless processing of petabyte-level datasets.
  • ​Simplified workflows: Intuitive APIs and minimal setup requirements.

​Getting Started with 3FS​

Step 1: Clone the Repository​

bash
git clone https://github.com/deepseek-ai/3fs  
cd 3fs  
git submodule update --init --recursive  
./patches/apply.sh

Step 2: Install Dependencies​
Supported on Ubuntu 20.04/22.04. Install prerequisites:

bash
# Ubuntu 20.04  
apt install cmake libuv1-dev liblz4-dev libboost1.71-all-dev ...

Step 3: Build and Deploy​

bash
cmake -S . -B build -DCMAKE_BUILD_TYPE=RelWithDebInfo  
cmake --build build -j 32

Full deployment guides are available in the 3FS documentation.

​Why Choose 3FS?

  • ​Open Source: Licensed under MIT, fostering community-driven innovation.
  • ​Cost Efficiency: Reduces reliance on expensive DRAM for caching.
  • ​Future-Proof: Optimized for next-gen AI models and hardware.

Explore Fire-Flyer File System (3FS)