Bobtail: Avoiding Long Tails in the Cloud
summary
Measurements show EC2 nodes exhibit persistent per-node long tails (99.9th percentile ~30 ms) caused by co-scheduling CPU-bound and latency-sensitive VMs. Controlled Zen testbed runs five VMs on two cores and reproduces order-of-magnitude tail increases when CPU-bound VMs meet core limits; a 30 ms scheduler timeslice explains the number. Bobtail runs a timer-interrupt oversleep test across k candidates, selects instances (40 chosen from 160) and yields up to 40% reduction in 99.9th-percentile latency while acknowledging throughput–latency tradeoffs.
AI-written summary · about
chapters (14)
- 0:00Start
- 0:20Public clouds are used everywhere
- 0:44Challenges to large applications
- 2:56Measurements Setup
- 4:28How does it compare to dedicated DCs?
- 5:07Spatial Properties
- 6:43Root cause analysis
- 9:25Why the type of sharing matters
- 13:19Evaluation
- 14:03Partition-Aggregation Workflows
- 17:31Conclusion
- 17:55Acknowledgements
- 19:27How bad is the latency tail?
- 21:47Discussion
related talks
21mPerformance Consistency in the Cloud
16m"Performance Testing: ..."
16mPreview: Cloud Computing
48mMeasuring performance variabillity of EC2
35mExtracting the Last Microseconds of Performance From Co...
36mMigrating Applications to the Cloud in a Cost Effective Manner
30mThinner Clouds with Preallocation
32mResearch Workloads on Preemptible VMs