Topic

Curvine: A Rust-Native Caching Filesystem for AI Workloads

Disaggregated storage has become the cloud-native standard, with object storage serving as the default data layer. However, AI workloads have made data movement efficiency a critical performance factor, from keeping GPUs fed during training and enabling KV cache reuse during inference to handling high-frequency, small I/O operations in AI agents. Simply adding a cache layer in front of object storage cannot overcome two inherent limitations: immutable objects lead to write amplification, while centralized NFS gateways become concurrency bottlenecks. Curvine takes a different approach: a Rust-native caching filesystem with full POSIX semantics (passing 1,129 LTP test cases), incremental synchronization through chunked writes, and parallel, direct data-plane access. It delivers approximately 100 μs metadata latency and supports up to 5 billion files per cluster, while allowing the cache to be discarded and data to remain recoverable. Four real-world deployments demonstrate measurable gains: a 23% reduction in both cross-cloud hot-table traffic and bandwidth consumption, 5.5× faster p50 latency for multimodal retrieval, a 2.7× improvement in peak time to first token (TTFT) through cross-Pod KV cache reuse, and zero Pending workspaces on Amazon EKS with just four core Pods supporting 10,000 workspaces. Outline: 1. Evaluation Criteria How to assess whether a distributed caching solution is suitable for AI training by examining three key factors: write amplification, gateway bottlenecks, and semantic gaps—without costly trial and error. 2. Architectural Trade-offs Why Rust is a good fit for storage systems, with no garbage collection, memory safety, and fearless concurrency; and how separating the control plane from the data plane enables four Pods to support 10,000 workspaces. 3. Reusable Performance Benchmarks Quantitative results—including a 23% reduction in hot-table traffic, 5.5× faster retrieval, 2.7× better peak TTFT, and six wins out of eight MLPerf AU benchmarks—that can be compared against in-house environments. 4. Technology Selection How to choose between FSMode and CacheMode based on mount points, and when to prioritize filesystem semantics versus speed, without treating them as mutually exclusive choices.