Storage Performance Testing in the Cloud
summary
Claims that public cloud storage performance is stable are false: repeated runs show 6,000–16,000 IOPS per instance and huge instance-to-instance variation, so capacity planning requires probability not averages. Practical guidance: measure latency distributions and the 99th percentile, avoid aggregation mistakes like stonewalling or summing thread results, and use real tools such as fio and filebench (IOzone, dbench, COSbench) with many runs, automation, and simulations to compute how many instances are needed for an SLA.
AI-written summary · about
chapters (17)
- 0:00Storage Performance Testing in the Cloud
- 0:26Overview
- 1:34Types of Performance
- 4:43Tail at Scale = Fail
- 6:19Aggregating Data
- 13:48Storage Performance Factors
- 16:28Data Testing: iozone
- 19:06iozone options
- 19:56Sample iozone output
- 20:45Metadata Testing
- 23:37Object Testing
- 25:44Noisy Neighbors
- 27:30Performance Distributions
- 30:23Performance Ratios
- 32:59Other Problems
- 36:26Conclusions
- 38:59Modeling Example
related talks
16mMore IOPS for Less: Exploiting Burstable Storage in Public Clouds
21mPerformance Consistency in the Cloud
37mCloud Storage 101
19mThree Actionable Ways to Improve your Cloud Storage
12mExponential Load Testing...
16m"Performance Testing: ..."
48mMeasuring performance variabillity of EC2
09mon What to Avoid in Public Clouds