UnICom: A Universally High-Performant I/O Completion Mechanism for Modern Computer...
summary
UnICom implements three kernel mechanisms—TagSched, TagPoll, and SKIP—in Linux to consolidate polling on a dedicated core, retain tagged sleeping tasks on CPU runqueues to avoid expensive dequeue/enqueue, and enable safe kernel-assisted direct SSD access that bypasses much of the I/O stack. Versus ext4, BypassD and io_uring it cuts 4 KB average latency by ~42% (single thread) and 128 KB latency by ~17.4%, improves P99 4 KB tail by ~31%, and matches or exceeds polling and interrupt trade-offs across low and high CPU utilization.
AI-written summary · about
chapters (15)
- 0:00UnICom introduction and motivation
- 1:43Interrupt vs polling dilemma
- 3:35Pure IO workload performance
- 6:03CPU contention in polling
- 8:30UnICom design goals
- 11:23IO latency breakdown insight
- 12:05UnICom architecture overview
- 14:09Tag schedule design
- 15:53Tech poll design
- 17:00Adaptive completion policy
- 17:43Experimental setup
- 18:04Pure IO results
- 19:32Mixed workload C# benchmark
- 20:34UnICom performance summary
- 21:17Conclusion and Q&A
related talks
59mThe Wild West of post-POSIX IO Interfaces
38mUh-oh, It's I/O Ordering!
11mUserspace Bypass: Accelerating Syscall-intensive Applications
21mPreview: Operating Systems
24mOptimizing Storage Performance for 4–5 Million IOPs
41mLinux Systems Performance
24mA Virtual IOMMU With Cooperative DMA Buffer Tracking
34mio_uring in the Automotive World