GPUs can DMA data directly from NVMe SSDs to keep the CPU off the data path, but doing so bypasses the OS file I/O layers and loses the page cache, read-ahead, and POSIX consistency. SPIN integrates peer-to-peer DMA into the standard OS file I/O stack, dynamically choosing between the P2P and page-cache paths, re-enabling read-ahead for sequential access while preserving POSIX semantics and portability across GPUs and SSDs. It exceeds raw P2P throughput by up to an order of magnitude and speeds real GPU-accelerated applications by 2.6–3.3×.

Published at USENIX ATC 2017 and in ACM Transactions on Computer Systems 2019.