λ-IO: A Unified IO Stack for Computational Storage
summary
λ-IO presents a full-stack software and hardware prototype (OpenSSD acting as NVMe) that routes per-request computation to whichever side is predicted faster, claiming up to 5.12× improvement. It implements an extended eBPF called sBPF runtime for safe in-device code, a lightweight dynamic dispatcher that profiles cache ratio and models transfer+compute cost, and evaluates on synthetic kernels and a modified Spark SQL to show when host or device wins.
AI-written summary · about
chapters (8)
- 0:00Intro
- 0:34Computational Storage
- 4:27A-10 APIs and Workflow
- 7:00Cross-Platform & Runtime
- 9:25Dynamic Request Dispatching
- 13:01Evaluation: Synthetic Application
- 14:50Evaluation: Case Study - Spark SQL
- 15:53Conclusion
related talks
19mAsynchronous I/O Stack: A Low-latency Kernel I/O Stack for Ultra-Low Latency SSDs
38mScaling IO-bound application in the cloud
22mCombining Buffered I/O and Direct I/O in Distributed File Systems
47mExploring Linux Memory Usage and IO Performance in Cloud Native Applications
40mAchieving the Ultimate Performance with KVM
30mOptimizing vLLM for Intel CPUs and XPUs
22mI/O in a Flash: Evolution of ONTAP to Low-Latency SSDs
01h 06mBest practices when accessing Big Data or any other data!