From Cloud to On-Prem - Capacity Planning and CPU Analysis
summary
Migration reverses a cloud cluster back on-prem for a large Aerospike cold store, using a POC of four AWS m5d metal machines versus four borrowed on-prem nodes. Measurement uses perf/flame graphs and Facebook psi plus Prometheus; a thread-safety bug forces remove-by-value to run 100% of writes, costing thousands of CPU cores worth of cycles and is fixed. Enabling ordered (k) maps saves ~16% CPU, turbo mode adds more, and combined changes make on‑prem ~30% faster and save over $1M.
AI-written summary · about
chapters (17)
- 0:00Introduction
- 1:25What is The Trade Desk
- 1:57Why Aerospike
- 3:59Setup Considerations
- 5:48System Performance
- 7:31Measuring CPU bottleneck
- 8:26Flame graphs
- 10:31Map add items
- 11:18Map data structure
- 13:52Key ordering
- 17:10CPU optimization
- 18:17Final build spec
- 19:15CPU and IO
- 20:09Benchmarking
- 23:31System of Record
- 24:42Life Path
- 28:49Conclusion
related talks
31mOn-Prem is the new Black
05mOn Prem is the new Black
25mBreaking Free From the Cloud: Banking on Self-Hosted Kubernetes
10mLightning Talk: Does Cloud Elasticity Pay off for HPC Workloads?
37mHybrid OnPrem + Cloud Architectures
17mOptimizing Cloud Compute Resources with Spare Cores
33mUncover Kubernetes cost-optimization trends
25mGreenOps at Scale: The 50TB Memory Diet for a Petabyte Platform