This 15-minute conference talk from USENIX FAST '25 explores the innovative approach of using on-device functions for data-dependent read I/O requests on read-only on-disk data structures. Learn how researchers from UNIST (Ulsan National Institute of Science and Technology) demonstrate that on-device I/O functions allow applications to issue requests more quickly while integrating with in-kernel functions to handle high request volumes efficiently. Discover the results of their prototype built on NVMeVirt, showing performance improvements including up to 41% throughput enhancement for B+-tree key-value stores and up to 3.85% reduction in WiredTiger's 99-percentile tail latency on YCSB C compared to host-only in-kernel storage functions.
Overview
Syllabus
FAST '25 - Selective On-Device Execution of Data-Dependent Read I/O
Taught by
USENIX