Past Experiences and Future Challenges using Automatic Performance Modelling to Complement Testing
- 1. Past Experiences and Future Challenges using Automatic Performance Modelling to Complement Testing Paul Brebner, CTO A NICTA/Data61/CSIRO Spin-out Company 16/03/2016 © Performance Assurance Pty Ltd 1
- 2. Performance modelling background • My background is analysis of distributed systems, middleware, GRID, architecture, performance, benchmarking (e.g. SPECjAppServer), sensor web performance, etc • Since 2007 project in NICTA to develop tools to assist mostly government systems of systems to perform better in advance • Service Oriented Performance Modelling tool • Model driven (SOA performance meta model) • GUI • Simulation for metric prediction • Enables modelling at level of workloads, composite and simple services, servers. • Used during early, middle, later lifecycle for lots of real systems 16/03/2016 Performance Assurance Pty Ltd 2
- 3. Performance modelling background • BUT Manual model building (structure, parameterisation, calibration) is • Time consuming • Expensive • Error prone • Limited to model complexity that can be built manually • Not easily repeatable or maintainable • Not accurate enough for some problems (need high quality and quantity of performance data) • Not fast enough for agile development • Last 3 years we have been a start up company, have to make $$$$$$ • Most customers have APM products • All customers want to increase speed and number of releases, reduce time and costs of testing • Solution is to use automatic model building from APM data • Cheaper and faster and more accurate • Solves new problems, e.g. DevOps 16/03/2016 Performance Assurance Pty Ltd 3
- 4. Automatic performance modelling from APM data • Only use available APM data • Use automatable (or potentially automatable) ways of getting the data from the APM into our Service Oriented Performance Modelling (SOPM) modelling/simulation tool (SaaS) • Automatically build and parameterise the performance data from the APM data • Multiple model types with various trade-offs, accuracy for capacity/response times, and model complexity/ability to change model aspects • Currently different model types are produced as part of the APM -> modelling tool transformation phase 16/03/2016 Performance Assurance Pty Ltd 4
- 5. Application DynatraceSF Dynatrace SF PurePath Dash Browser PP XML Converter Model XML Modelling SaaS 1 2 3 4 5 SF Dynatrace Session File PP XML Dynatrace Server REST API PurePath XML File Model XML XML Model File KEY
- 6. Dynatrace Transaction flow dashboard 16/03/2016 Performance Assurance Pty Ltd 6
- 7. Produces: Simple capacity model 16/03/2016 Performance Assurance Pty Ltd 7
- 8. Dynatrace PurePath Dashboard (detailed per transaction call tree) 16/03/2016 Performance Assurance Pty Ltd 8
- 9. Produces: Transactional model (portion) 16/03/2016 Performance Assurance Pty Ltd 9
- 10. Experiences with three projects • Project 1 • P2V migration • Project 2 • C2V test -> prod • Project 3 • DevOps • Focus of this talk, come to main ICPE talk for others 16/03/2016 Performance Assurance Pty Ltd 10
- 11. Project 3 • Devops • Focus on response time SLAs • Deployment/resources • Faster cycle time • More releases • Less and cheaper testing • Challenge • Proprietary in-house APM tool • “Profile point” times only • Required pre-processing (using Hive) 16/03/2016 Performance Assurance Pty Ltd 11
- 12. Focus • Risk service • Heavily used • Multiple services • New services added all the time • Services had different time and memory profiles • Would a new service break the SLA? • Baseline model accurate to 10% response time 16/03/2016 Performance Assurance Pty Ltd 12
- 13. Alternatives modelled • Changing transaction mix • Changing arrival rates • Making some services asynchronous, concurrent • Adding new risk assessment services • More complex • Optimising deployment of services to multiple servers taking into account memory and CPU usage, and response time • A type of box/bin packing problem • 4 services out of 30 used 50% of CPU 16/03/2016 Performance Assurance Pty Ltd 13
- 14. Challenges • Pre-processing APM data “profile points” • Low load for APM data sample c.f. target load • Used calibration from load tests on pre-production to improve accuracy • No CPU time breakdown from APM data • But GC had a profile point (and was significant) • Transaction types not in APM data • Had to infer them, either too few or too many 16/03/2016 Performance Assurance Pty Ltd 14
- 15. 16/03/2016 Performance Assurance Pty Ltd 15
- 16. 16/03/2016 Performance Assurance Pty Ltd 16
- 17. 16/03/2016 Performance Assurance Pty Ltd 17
- 18. DevOps • Goal is to shift left and shift right • Shift right • Build and continuously maintain performance model of production to accurately model response times, scalability, capacity and resource requirements under target production loads • Shift left • Calibrate production performance model for development • Enable developers to make code changes, explore impact with unit tests and development APM to incrementally rebuild performance models • To understand likely performance and scalability impact • Speed up development cycle as no longer have to wait (weeks) for performance testing 16/03/2016 Performance Assurance Pty Ltd 18
- 19. Existing Dev, Test, Prod lifecycle: Delays in feedback: Takes weeks per iteration, test env is a bottleneck, environments are different Dev Test Prod Late Feedback Late Feedback Deploy to test Deploy to prod
- 20. DevOps + APM: earlier but not completely accurate performance feedback i.e. environments are different so APM data is different across lifecycle Dev Test Prod Late Feedback Deploy to test Deploy to prod APM APM APM Earlier Feedback
- 21. DevOps + APM + Modelling: Earlier more accurate performance predictions -> decreased cycle time 16/03/2016 Performance Assurance Pty Ltd 21 Dev Test Prod Deploy to test Deploy to prod APM APM APM Early Feedback Base Model Dev Model Incremental updates to Base model with dev changes Baseline model buildDev Model Update Calibrate prod model for dev
- 22. Benefits • Changes in code in Dev • Unit test • APM performance data • Incrementally update calibrated performance model • Predict performance and scalability impact for Prod env • Cheaper and faster than waiting for testing and deployment to Prod • Sensitivity analysis could determine areas of greater sensitivity to changes and thresholds • These would be subject to more rigorous modelling and testing 16/03/2016 Performance Assurance Pty Ltd 22
- 23. DevOps + APM + Modelling: In reality lots of dev, different environments 16/03/2016 Performance Assurance Pty Ltd 23 Dev Test Prod Deploy to prod APM APM APM Base Model Dev Model Baseline model build Dev APM Dev Model Dev APM Dev Model Dev APM Dev Model
- 24. Challenges • Calibration of performance models for use in Dev from Test and Prod • Once predictions are made how do we test if they are supported by the APM data or not? i.e. if null hypothesis is “changes in dev will have no impact on prod”, how do we determine if this is supported by evidence or not? • Is it scalable? • Lots of developers and changes to subsets of code • Concurrent and compounding changes would need centralised model with all changes incorporated • What about changes to infrastructure code that could impact everything? • How to support this in Dev APM and modelling tools • ROI • Depending on cost of testing, cost of initial setting up APM and modelling tools and incremental costs, number of tests and modelling predictions per cycle, and value of reduced cycle times and earlier performance predictions, ROI may occur earlier or later or never… • Example • Assumes model calibrated once per release from performance APM data • Assumes one actual load test per release • What’s tradeoff between multiple tests per release vs 1 test and multiple modelling predictions? 16/03/2016 Performance Assurance Pty Ltd 24
- 25. Costs: Modelling cheaper after 3 changes 16/03/2016 Performance Assurance Pty Ltd 25 0 5000 10000 15000 20000 25000 30000 35000 40000 45000 0 2 4 6 8 10 12 Cost($) Number of changes tested/modelled Costs of LoadTest only and HybridApproach LoadTestOnly Modelling
- 26. Speed: Average hours to test/model a number of code changes (per model calibration) 16/03/2016 Performance Assurance Pty Ltd 26 0 10 20 30 40 50 60 70 80 90 0 2 4 6 8 10 12 Averagetime(hours) Number of changes per calibrated model Average hours to test/model changes LoadTestAvgHourPerChange ModellingAvgHoursPerChange
- 27. Send us your data • Free trial of simple Dynatrace capacity models • http://www.performance-assurance.com.au/send-us-your-data/ • http://www.performance-assurance.com.au/introduction-to- automatic-model-building/ • Send us a sample Dyntrace session file and we’ll send you a link to a demo capacity model • Particularly interested in trending technologies and use cases, e.g. Micro-services, Containers, Big Data, IoT, etc • Free Personal Dynatrace license from: http://bit.ly/dtpersonal 16/03/2016 Performance Assurance Pty Ltd 29