Shinjuku: Preemptive Scheduling for μsecond-scale Tail Latency

on YouTube ↗24m

" summary

Shinjuku is a single-address-space OS that uses hardware virtualization features to enable frequent preemption (as often as every 5 μs) so centralized schedulers can avoid short requests being blocked by long ones. It targets μs-scale dataplanes like IX and ZygOS and shows improved tail latency and throughput—e.g., up to 6.6× throughput and 88% lower tail latency on a RocksDB mixed point/range workload—especially for workloads with high-dispersion or heavy-tailed service times.

AI-written summary · about

" chapters (15)

1. 0:00Start
2. 0:29Tail latency matters for datacenter workloads
3. 1:07Achieving low tall latency of microsecond scale is hard
4. 3:08Ideal centralized queue is better in simulation
5. 4:47Problem: Short requests get stuck behind long ones
6. 6:45Solution: Shinjuku
7. 7:10Outline
8. 9:07Minimizing Preemption Overhead
9. 12:09Scheduling policy
10. 13:14Queue Selection Policy
11. 14:34Evaluation
12. 15:34Shinjuku under low variability
13. 16:33Shinjuku under high variability
14. 18:17Does Shinjuku scale?
15. 19:17Conclusion