Vigil-KV: Hardware-Software Co-Design to Integrate Strong Latency Determinism...

on YouTube ↗20m

" summary

Claims that predictable latency mode (PLM) per NVMe partition and coordinated host scheduling eliminate long-tail reads, implemented by partitioning a 1.92TB NVMe SSD prototype into multiple kvset/meta set physical functions and adding IO-priority hooks into LSM KV (Linux 4.19.91). Device-state scheduling hides SSD non-deterministic windows while compaction/flush and client requests are co-scheduled; evaluation on Facebook/Yahoo workloads shows about 3.19x tail reduction and 34% average latency cut.

AI-written summary · about

" chapters (24)

1. 0:00Intro
2. 0:40Goal: Guarantee Strong Latency Determinism
3. 1:31Service Example: Facebook
4. 1:55Service Latency Breakdown
5. 2:41Overview of Backend
6. 4:17Impacts of LSM KV's Internal Tasks
7. 5:48Limitation: Internal Task Scheduling
8. 7:44Existing Solution: NVMe's PLM Interface
9. 9:47Opportunity of Space-Division
10. 10:03Opportunity of Time-Division (= PLM I/F)
11. 10:33PLM Constraints
12. 10:44Write-free on DTWIN
13. 11:16Fair-scheduling for PLM Winc
14. 12:33Vigil-KV Software
15. 13:25Device State Scheduling
16. 14:54Request Scheduling
17. 15:25Vigil-KV Hardware Prototype
18. 15:54#1 Partitioning an SSD
19. 16:27Aside: Performance Validation
20. 17:19#2 Integrating PLM
21. 18:06PLM Constraint & Behavior Analysis
22. 18:45Device Lockdown Constraint
23. 19:02Summary of Revealed PLM Behaviors
24. 19:14Representative Evaluation Result