eurostar 2013 albert witteveen final
TRANSCRIPT
With cloud computing:who needs performance testing
Albert Witteveen
You just woke up after a 10 years nap:
Team member:“We can add extra processing power and memory on the fly.
An extra database has a lead time of two weeks”
Imagine
Does this sound familiar: Performance test: everything OK Day 1 on production: we end up adding more than four
times the hardware
Question
1. the tools simulate but are not quite equal2. load profiles are based on too many assumptions3. we report more accurately than we can measure4. long setup time → limited amount of tests5. we hide it all in complex reports
Load testing weak points
We send and accept the same requests and responses but can't anticipate slight changes
In production, a lot more is going on than just our test Did we really get a good response Similar hardware is expensive
Our tools simulate reality but are not equal
Cloud computing: adding extra hardware can be done on the fly and on a moments notice
With the high costs of performance testing and how easy we can 'speed things up' if needed:
Why bother testing? The money is better spent on that extra hardware
Cloud computing
Just start with an overkill of hardware and scale down to what is actually used!
Dutch auction?
Introducing: The Queuing theory
The naïve tester (me)
Computers are running or idling. The queuing theory is an established model for
performance engineers It can describe the behavior of systems on every layer
Queuing theory
Simple queue
Multiple registers
Take a number
Queuing center: a location in our system where waiting (queuing) occurs a Bottleneck if you will
◦ They can exist anywhere: CPU, Memory, Network, IO, other systems◦ There is always one or more queuing centers◦ A queuing center really determines the performance◦ The queuing center provide key information on scalability◦ Service and wait time are the real components of performance
Queuing center
Queuing model describe anything: large connected systems, small, embedded ...
You can 'zoom in' and the model can describe the behavior or the server
You can keep zooming in to CPU, network etc.
Multiple layers
Multiple zoom levels Residence time = wait + service time There is always a queuing center No queuing center found: look harder
Queuing models
Cloud computing not infinite: Financial limit Technical: IO/Network/CPU speed per process
We don't build supercomputers to calculate a mortgage offer
Back to the cloud
Always find the queuing centers Based on the result: judge 'yes we are likely to meet
requirement X Y and Z' Show where the risks are 'requirement x cannot be feasibly
met for function y' Explore the risks
How to apply the queuing theory
Explore identified resource heavy components with stakeholders, developers and oracles◦ Other use of this component?◦ Real frequency of usage?◦ Validity of the (generic) requirement for this function?
Place the results in context: ◦ You may have a bigger issue than you thought◦ Or it is actually OK for this usage
Exploring the risks
Define a set of key functions/use cases with stakeholders and experts (i.e. functional testers)
Per test identify at least one queuing center Compare with generic requirements
◦ Can meet ?◦ Risk exist → explore → place in context →define further test
The model allows you to place real behavior in context and a realistic assessment of risk
Approach
If no queuing center was found → monitoring was not sufficient
Queuing centers:◦ Tell you about the risks to core functionality: performance and
financial◦ Tell you on the ability to scale◦ Improve response time in scaling up
For stakeholders
Stakeholders don't (necessarily) understand queuing models
Explain in what matters to them: i.e. when making the offer it takes 15 seconds to generate
Think of the systems as queuing systems and explain behavior
Make a model?
Knowing what the behavior is can tell you:◦ if you can handle requirements ◦ how to scale if needed◦ estimate if performance can be met within budget◦ if you need to adapt your cloud (i.e. improve IO/network, CPU)
So yes: it still makes sense to do performance testing
Summary for the cloud
Batch process tested to be run from multiple servers Process needed to be faster Risk: 'on-line' processes on server should not be impacted
Finding: 3 servers, three times as fast. But no queuing center found???
Deep diving in CPU monitoring showed the queuing center: Process was pausing/waiting after each cycle
Conclusion: → on-line processes not impacted as there was sufficient CPU time for other processes
Example: batch process
Stress point found Unclear where queuing center was
Cause: JAVA memory management can be deceiving on OS level.
Rule that the queuing center needed to be found made us find out. The absence of a queuing center makes you look further
Example JAVA
Next