Telescope: Telemetry for Gargantuan Memory Footprint Applications
summary
Telescope is a kernel telemetry technique that profiles higher and lower levels of the radix page table to find hot and cold pages in terabyte-class address spaces, implemented as a region-based sampler that picks the highest-covering PTE when possible. On a 5 TB microbenchmark it achieves 90%+ precision and recall while costing 0.9% single CPU utilization, finds 10–30 GB of hot working set where Damon finds nothing, and yields 5.6%–34% throughput improvement on 1–2 TB real workloads versus PEBS and Damon.
AI-written summary · about
chapters (7)
- 0:00Telescope introduction
- 0:43Tiered memory systems
- 3:12Prior telemetry limits
- 8:51Telescope design
- 15:31Evaluation vs Damon and PEBS
- 18:02End-to-end performance
- 19:25Conclusion
related talks
24mScaling Telemetry To Terabytes: Applying Google's Principles To OpenTelemetry...
15mTuning OTel Collector Performance Through Profiling
25mPerfect Match: Correlating Continuous Profiling with Distributed Tracing for Stronger Observability
21mOpen Telemetry
24mUnlocking Advanced Processing Capabilities with OpenTelemetry
16mBPF Based Telemetry, Metrics, and Accounting on Android
31mTroubleshooting and Monitoring Production Applications using OpenTelemetry
27mThe Signal in the Storm: Practical Strategies for Managing Telemetry Overload