3FS

Questions & Answers

What is 3FS (Fire-Flyer File System)?

The Fire-Flyer File System (3FS) is a high-performance distributed file system developed by DeepSeek AI. It is engineered to handle the intensive storage demands of modern AI training and inference workloads. It provides a shared storage layer, utilizing SSDs and RDMA networks for optimal performance.

Who is 3FS designed for?

3FS is designed for developers and organizations working with large-scale AI training and inference applications. It targets users who require a high-throughput, low-latency storage solution for tasks like data preparation, dataloading, checkpointing, and KVCache for LLM inference.

What makes 3FS different from other distributed file systems?

3FS distinguishes itself through a disaggregated architecture that combines throughput from thousands of SSDs and network bandwidth from hundreds of storage nodes, making storage access locality-oblivious. It also implements Chain Replication with Apportioned Queries (CRAQ) for strong consistency, simplifying application development.

In what scenarios should 3FS be used?

3FS should be used in environments requiring high-performance storage for AI workloads, such as organizing outputs of data analytics pipelines, enabling random access for deep learning dataloaders, supporting high-throughput parallel checkpointing for large models, and providing cost-effective KVCache for LLM inference.

What is the peak read throughput achieved by 3FS?

In a read stress test with 180 storage nodes and approximately 500 client nodes, 3FS demonstrated an aggregate read throughput of approximately 6.6 TiB/s. Each storage node was equipped with 2x200Gbps InfiniBand NICs and sixteen 14TiB NVMe SSDs.