Optimized Scatter/Gather Support for Parallel Storage
summary
A patch implements end-to-end vectored I/O (readv/writev and io_uring-style) through a parallel storage stack targeting Ceph OSD, eliminating intermediate memcpy by passing iovecs from client to storage. The talk covers code diffs, a test harness, microbenchmarks with latency and throughput graphs, and a fallback path when kernel support is missing; results show improved steady-state efficiency but require kernel features for full zero-copy.
AI-written summary · about