Priority IO Scheduling in the Cloud
summary
Implements priority-based (PBS) IO scheduling by passing priorities from HDFS through a user-level JNI + SG_IO path into the Linux kernel and SATA device using ICC NCQ high-priority commands, focused on read paths and bypassing the page cache. Tested on a 4‑node cluster (four HDDs/node) with YCSB and MapReduce contention; PBS keeps latency low while allowing higher queue depth, showing up to ~5x improvement in worst‑case latency and >2x improvement in average latency versus Linux CFQ, at the cost of added stack changes and potential starvation of low‑priority IO.
AI-written summary · about