OffloadFS: Leveraging Disaggregated Storage for Computation Offloading
New file system taps into idle compute power in storage hardware, accelerating RocksDB and ML prep by up to 3.36x.
A research team from KAIST has introduced OffloadFS, a novel user-level file system designed to unlock the latent compute and memory capacity within modern disaggregated storage infrastructure. In data centers using NVMe over fabrics (NVMeoF), storage resources are separated from compute nodes, but the storage hardware itself often sits idle beyond basic I/O tasks. OffloadFS enables near-data processing by safely offloading specific, I/O-intensive workloads directly to these storage nodes or to peer compute nodes, bypassing the overhead of traditional distributed lock management. This approach optimizes cache usage by reducing interference between threads performing different operations.
Built on top of OffloadFS, the team created two practical applications: OffloadDB and OffloadPrep. OffloadDB allows the popular RocksDB key-value store to offload critical but resource-heavy operations like MemTable flushing and compaction to storage hardware. OffloadPrep similarly offloads image pre-processing tasks for machine learning pipelines. In evaluations, this architecture delivered substantial performance gains, speeding up RocksDB operations by up to 3.36x and ML pre-processing tasks by 1.85x compared to the conventional OCFS2 file system. The work, submitted to IEEE, represents a significant step towards more efficient data center resource utilization by treating smart storage as a computational resource rather than just a passive data repository.
- OffloadFS taps into idle compute in NVMe-over-fabrics storage nodes for near-data processing, improving RocksDB performance by 3.36x.
- The system eliminates the need for distributed lock management and reduces cache interference between I/O threads.
- Two built applications, OffloadDB and OffloadPrep, enable offloading of database compaction and ML image pre-processing tasks directly to storage hardware.
Why It Matters
This could dramatically reduce latency and cost for data-intensive applications like real-time databases and AI training pipelines in cloud data centers.