Measuring performance variabillity of EC2

on YouTube ↗48m

" summary

MongoDB performance tests run 25 times each to measure variability and the big findings are specific: disk noise from ephemeral SSDs is worse than EBS with provisioned IOPS so switching to 5,500 IOPS EBS cuts variance, microbenchmarks with FIO and network checks with iperf3 isolate causes, and CPU tuning — turning off hyperthreading plus disabling power-save states and using numactl — reduces unpredictability. Results drop most tests below 10% variance and many under 5% so the changes are deployed to production.

AI-written summary · about

" chapters (9)

1. 0:00Start
2. 3:31Performance Testing
3. 7:56Placement Groups
4. 29:29Cpu
5. 30:23Hyper Threading
6. 32:43Mongodb Results
7. 35:40Network Tests
8. 37:33Canary Tests
9. 40:45What Exactly Was Wrong with Placement Groups

" related talks

• 21mPerformance Consistency in the CloudNati Cohen, Sercan Karaoğlu · SREcon25 Europe/Middle East/Africa
• 42mStorage Performance Testing in the CloudJeff Darcy · LISA '13
• 16m"Performance Testing: ..."Imad Mouline · Velocity 2010
• 25mGreenOps at Scale: The 50TB Memory Diet for a Petabyte PlatformLeonardo Santos · DevOpsDays Geneva 2026
• 22mOPTIMUSCLOUD: Heterogeneous Configuration Optimization for Distributed DatabasesSaurabh Bagchi, Ashraf Mahgoub, Subrata Mitra, Somali Chaterji, Ana Klimovic, Alexander Michaelson Medoff, Rakesh Kumar · USENIX ATC '20
• 30mAccelerating innovation with Amazon EBSAshish Palekar · AWS AMER Summit May 2021
• 18mWhat's New with Amazon Elastic Block Store (Amazon EBS)Ashish Palekar · AWS Storage Day 2020
• 12mExponential Load Testing...Saurabh Bajaj · Velocity Conference Santa Clara 2014