J
JonLester
Benchmark performance testing involves measuring the performance characteristics of an application or system under normal or expected conditions. It's a recommended practice in any case, but it's a critical consideration for your APIs since your consumers will depend on consistent performance for their client applications.
Incorporating benchmark testing of your Microsoft Azure API Management services into your software delivery process provides several important benefits:
While the approaches and tools involved are nominally very similar, the reasons for doing them differ. Benchmark testing establishes a performance baseline within the normal operational range of conditions, while load testing established the upper boundary or point of failure. Benchmark testing establishes a reference point for future iterations, while load testing validates scalability and stress handling. Both are important for ensuring API performance, and you can combine the approaches to suit your needs as long as the goals of each are met.
Below, we'll describe the principles of designing a repeatable benchmark test and conclude with a full walkthrough and the resources you'll need to do it yourself.
Before we get into a specific example, let's look at the conceptual steps involved.
Broadly, there are two stages:
The execution stage is repetitive. The first execution result becomes the baseline. From there, the benchmark test can be repeated after any important change to your API workload (resource configuration, backend application code, etc.). Comparing the results of the current and previous test will indicate whether the most recent change moved you closer to your goal or caused a regression. Once the goal is met, you'll continue the practice with future changes to ensure that the required performance is being maintained.
Determine the key performance metric that will define your benchmark. Think of it as the key performance indicator (KPI) of your API workload. Some examples include: operation or request duration, failure rate, resource utilization (eg, memory usage), data transfer speed, and database transaction time. The metric should align with your requirements and objectives, and be a good indicator for the quality of the consumer experience. For API Management, and APIs in general, the easiest and most useful metric is usually response time. For that reason, start with response time as the default choice if your circumstances don't guide you to choose something else.
The key here is to choose a single metric that you can capture easily and consistently, is an indicator of the kind of performance you are after, and that will allow you to make linear comparisons over time. It's possible to devise your own composite metric based on an aggregation formula using multiple primitives, if required, in order to derive a single benchmark measurement that works best for you.
The scenario describes input parameters and the simulation. In other words, it describes what is happening in the system while the benchmark metric is being measured. For example, "1000 simulated users, calling the Product Search API, at a rate of 10 searches per minute per user". The scenario should be as simple as possible while also providing a realistic representation of typical usage and conditions. It should accurately reflect the behavior of the system in terms of user interactions, data payloads, etc. For example, if your API relies on caching to boost performance, don't use a scenario that results in an unrealistically high cache hit rate.
The test environment includes the tool that will run the simulation (JMeter, for example), along with all the resources your API requires. Generally speaking, you should use a dedicated environment that models your production environment as closely as possible, including compute, storage, networking, and downstream dependencies. If you have to use mocks for any of your dependencies, make sure that they are accurately simulating the real dependency (network latency, long running processes, data transfer, etc).
You may need to instrument your code or API Management service with performance monitoring tools or profiling agents (for example, Azure Application Insights). You may also need to consider how you will retrieve and store the results for future analysis.
Run the defined test scenario against the API while measuring the performance metric.
Analyze the collected performance data to assess how your API performs. If this isn't your first time running the test, compare the observed performance against previous executions to determine if the API continues to meet the desired performance objectives and what the impact (if any) of your code or configuration changes may be. There are statistical methods that can applied to aid this analysis, which are extremely useful in automated tests or pull request reviews. These methods are beyond the scope of this post, but it's a good idea to familiarize yourself with some of the approaches.
Document the test results, including performance metrics, observations, and any identified issues or recommended actions. This information serves as a reference for future performance testing iterations and as a new benchmark for future comparison.
Finally, find ways to automate or optimize the process or modify your strategy as necessary to improve its usefulness to your business operations and decision making. In a future article, we'll talk more about how to operationalize benchmark testing and how to use it as a powerful capacity management tool.
Let's make this more realistic with a basic example. For the purposes of this walkthrough, we've developed an automated environment setup using Terraform. Find more information about the environment and the source code on GitHub. The environment includes an API Management service, a basic backend (httpbin, hosted in an Azure App Service plan), and an Azure Load Testing resource.
You are free to choose any testing tools that fit your needs, but we recommend Azure Load Testing. It doesn't require you to install JMeter locally or author your own test scripts. It allows you to define parameters, automatically generates the JMeter script for your test, and manages all the underlying resources required for the test agents. Most importantly, it avoids many of the problems we'd be likely to encounter with client-based tools and gives us the repeatability we need.
Let's look at how we'll apply our model approach in the example:
1. Open Azure Cloud Shell and run the following commands.
2. Clone the Repository
3. Initialize Terraform
4. Plan the Deployment
5. Apply the Terraform Templates
In our first benchmark, we were establishing a performance baseline of the backend application which returns a 500-byte payload. We tested the backend in isolation (meaning the load test client was sending requests directly to the backend API end point, without API Management) so that we could measure how it performs on its own. This isn't always necessary, or even practical, but it can provide really useful insights. Below are the results from three different runs of that first test:
First result set
Next, we ran the same benchmark test using the API Management endpoint so requests were being proxied through API Management to the backend application. This scenario is an "end-to-end" or "system" test that is representative of how our API would be deployed in production. The results help us measure any latency or change in performance added by API Management and the Azure network. As we can see, the results are similar. This indicates that the net effect of API Management on the system performance at this design load is zero or very close to zero.
Second result set
Finally, we ran a benchmark on a new "release" of our backend application. The new version of the API now returns a larger 1,500 byte payload, and we can see from the results that response times have increased significantly.
Third result set
Assuming these results don't meet our performance objectives, we now know that remediation steps will need to be taken before the new release of our API can be deployed to production. For example, we might consider adding output caching, or scaling the App Service or API Management service, or look for ways to optimize the payload returned from the application code. In any case, we now have the tools to test any remediation approach (using the same structured, quantitative approach above) so that we can be sure that the new API version meets its performance objective before it's released.
Continue reading...
Incorporating benchmark testing of your Microsoft Azure API Management services into your software delivery process provides several important benefits:
- It establishes performance baseline as a known, quantifiable starting point against which future results can be compared.
- It identifies performance regressions so that you can pinpoint changes or integrations that may be causing performance degradation or hindering scalability — in effect helping you to identify which components might need to be scaled or configured to maintain performance. This allows developers and operational staff to make targeted improvements to enhance the performance of your APIs and avoid accumulating performance-hindering technical debt.
- It validates performance requirements so you can be assured that the architecture meets the desired operating performance targets. This can also help you determine a strategy for implementing throttling or a circuit breaker pattern.
- It improves user experience by identifying and resolving performance issues early in the development life cycle— before your changes make it into production.
- And perhaps most importantly, it gives you the data you need to create the capacity model you'll need to operate your APIs efficiently across the entire range of design loads. This is a topic for a future post, but the methods described here are a a great starting point.
Benchmark vs Load Testing. What's the difference?
While the approaches and tools involved are nominally very similar, the reasons for doing them differ. Benchmark testing establishes a performance baseline within the normal operational range of conditions, while load testing established the upper boundary or point of failure. Benchmark testing establishes a reference point for future iterations, while load testing validates scalability and stress handling. Both are important for ensuring API performance, and you can combine the approaches to suit your needs as long as the goals of each are met.
Below, we'll describe the principles of designing a repeatable benchmark test and conclude with a full walkthrough and the resources you'll need to do it yourself.
A model approach
Before we get into a specific example, let's look at the conceptual steps involved.
Broadly, there are two stages:
- Design and Planning: Decide what to measure, and how to measure it. (Steps 1-4 below)
- Execution: Run test, collect results, and use the results to inform future actions or decisions. (Steps 5-7)
The execution stage is repetitive. The first execution result becomes the baseline. From there, the benchmark test can be repeated after any important change to your API workload (resource configuration, backend application code, etc.). Comparing the results of the current and previous test will indicate whether the most recent change moved you closer to your goal or caused a regression. Once the goal is met, you'll continue the practice with future changes to ensure that the required performance is being maintained.
1. Identify your benchmark metric
Determine the key performance metric that will define your benchmark. Think of it as the key performance indicator (KPI) of your API workload. Some examples include: operation or request duration, failure rate, resource utilization (eg, memory usage), data transfer speed, and database transaction time. The metric should align with your requirements and objectives, and be a good indicator for the quality of the consumer experience. For API Management, and APIs in general, the easiest and most useful metric is usually response time. For that reason, start with response time as the default choice if your circumstances don't guide you to choose something else.
The key here is to choose a single metric that you can capture easily and consistently, is an indicator of the kind of performance you are after, and that will allow you to make linear comparisons over time. It's possible to devise your own composite metric based on an aggregation formula using multiple primitives, if required, in order to derive a single benchmark measurement that works best for you.
Tip: Requests per second (RPS) might be the first metric you think of when you are trying to decide what you should measure. Similar unit-per-second metrics have been used historically for benchmark everything from web servers to GPUs. But in reality, RPS by itself isn't very useful as a benchmark for APIs. It's not uncommon to observe a system achieve a "high" RPS while individual consumers are simultaneously experiencing "slow" response times. For this reason, we recommend that you only use RPS as a scenario parameter and choose something else as your benchmark metric.
2. Define the benchmark scenario
The scenario describes input parameters and the simulation. In other words, it describes what is happening in the system while the benchmark metric is being measured. For example, "1000 simulated users, calling the Product Search API, at a rate of 10 searches per minute per user". The scenario should be as simple as possible while also providing a realistic representation of typical usage and conditions. It should accurately reflect the behavior of the system in terms of user interactions, data payloads, etc. For example, if your API relies on caching to boost performance, don't use a scenario that results in an unrealistically high cache hit rate.
Tip: For an existing application, choose an API operation that represents an important use case and is frequently used by your API consumers. Also, make sure that the performance of the scenario is relatively deterministic— meaning that you expect the test results to be relatively consistent across repeated runs using the same code and configuration, and the results aren't likely to be skewed by external or transient conditions. For example, if your API relies on a shared resource (like a database), make sure the external load on that resource isn't interfering with your benchmark. When it doubt, use multiple test runs and compare the results.
3. Define the test environment
The test environment includes the tool that will run the simulation (JMeter, for example), along with all the resources your API requires. Generally speaking, you should use a dedicated environment that models your production environment as closely as possible, including compute, storage, networking, and downstream dependencies. If you have to use mocks for any of your dependencies, make sure that they are accurately simulating the real dependency (network latency, long running processes, data transfer, etc).
Tip: You want your testing environment to satisfy two conditions:
- It makes it easy to set up and execute the test. You don't want to deter yourself from running tests because the process is tedious or time-consuming.
- It is consistent and repeatable across test runs to ensure the observed results can be compared reliably.
Automation helps you achieve both of these things.
4. Determine how you will record your chosen metric
You may need to instrument your code or API Management service with performance monitoring tools or profiling agents (for example, Azure Application Insights). You may also need to consider how you will retrieve and store the results for future analysis.
Tip: Be aware that adding observability and instrumentation can, by itself, adversely impact your performance metric, so the ideal case (if the observability tooling isn't already part of your production-ready design) would be a data collection method that captures the data at the client (or agent, in the case of Azure Load Testing).
5. Execute the test scenario
Run the defined test scenario against the API while measuring the performance metric.
6. Analyze the results
Analyze the collected performance data to assess how your API performs. If this isn't your first time running the test, compare the observed performance against previous executions to determine if the API continues to meet the desired performance objectives and what the impact (if any) of your code or configuration changes may be. There are statistical methods that can applied to aid this analysis, which are extremely useful in automated tests or pull request reviews. These methods are beyond the scope of this post, but it's a good idea to familiarize yourself with some of the approaches.
For Example: You just added a policy change that decrypts part of the request payload and transforms it into a different format for your backend to consume. You noticed that the time for the operation to complete has increased from 70ms to 110ms. Your benchmark objective is 80ms. Do you revert the change? Do you scale your API management service to compensate? Do you try to optimize your recent changes to see if you can get the results to improve? The bottom line here is that you can use the data to make an informed decision.
7. Report and document
Document the test results, including performance metrics, observations, and any identified issues or recommended actions. This information serves as a reference for future performance testing iterations and as a new benchmark for future comparison.
8. Iterate and refine
Finally, find ways to automate or optimize the process or modify your strategy as necessary to improve its usefulness to your business operations and decision making. In a future article, we'll talk more about how to operationalize benchmark testing and how to use it as a powerful capacity management tool.
Walkthrough
Let's make this more realistic with a basic example. For the purposes of this walkthrough, we've developed an automated environment setup using Terraform. Find more information about the environment and the source code on GitHub. The environment includes an API Management service, a basic backend (httpbin, hosted in an Azure App Service plan), and an Azure Load Testing resource.
Tip: Use the Terraform templates provided in the repo to deploy all the resources you'll need to follow along. For operational use, we recommend that you create your own repository using our repo as a template, and then follow the instructions in the README to configure the GitHub workflows for deployment to your Azure subscription. Once configured, the workflow will deploy the infrastructure and then run the load tests for you automatically.
You are free to choose any testing tools that fit your needs, but we recommend Azure Load Testing. It doesn't require you to install JMeter locally or author your own test scripts. It allows you to define parameters, automatically generates the JMeter script for your test, and manages all the underlying resources required for the test agents. Most importantly, it avoids many of the problems we'd be likely to encounter with client-based tools and gives us the repeatability we need.
Let's look at how we'll apply our model approach in the example:
Performance metric | Average response time |
Benchmark scenario | Performance will be measured under a consistent request rate of 500 requests per second. |
Environment | The sample environment - an App Service Web App that hosts the backend API and an API Management Service configured with one scale unit. Both are located in the same region, along with the Azure Load Testing resource. The deployment assets for all resources are included. |
Deploy the Azure resources
1. Open Azure Cloud Shell and run the following commands.
2. Clone the Repository
Code:
git clone https://github.com/ibersanoMS/api-management-benchmarking-sample.git
cd api-management-benchmarking-sample/src/infra
3. Initialize Terraform
Code:
terraform init
4. Plan the Deployment
Code:
terraform plan -out=tfplan
5. Apply the Terraform Templates
Code:
terraform apply tfplan
Creating and running the tests
Note: The Terraform templates will configure the load tests for you, but if you want to create tests on your own the steps below will walk you through it.
Identify the host url of your App Service backend and your API Management service. If you're using the sample environment created from the Terraform template, these will be the "backendUrl" and "apiUrl" respectively.
Search for Azure Load Testing in the Azure Portal.
- Click Create on the resource provider menu bar.
- Once the Load Testing resource is created, navigate to Tests.
Click Create on grid menu bar and then choose Create a URL-based test.
Create a new test
Configure the test with the following parameters for your first case (500B payload). Enter the App Service backend as the host portion of the Test Url, which should be in the form of:https://{your App Service hostname}/bytes/500
.
Configure the test
Click Run Test. Once the test completes, you should see results like below:
View the results
Now that we have a baseline result for the backend, create and run another identical test, but this time use the API Management API URL as the Test Url (https://{your API Management service hostname}/bytes/500
).
Finally, we'll simulate an updated version of the API by increasing the response payload size. Our API now returns more data than the previous version, so we'll be able to measure the impact of that change.
Configure and run a new test. We're still using the API Management host url, with a url path that returns 1500 bytes instead of 500 bytes: (https://{your API Management service hostname}/bytes/1500
).
Create a test simulating an API change
Once the test completes, you should see results like below:
Results of the new test
Looking at the Results
In our first benchmark, we were establishing a performance baseline of the backend application which returns a 500-byte payload. We tested the backend in isolation (meaning the load test client was sending requests directly to the backend API end point, without API Management) so that we could measure how it performs on its own. This isn't always necessary, or even practical, but it can provide really useful insights. Below are the results from three different runs of that first test:
Throughput (RPS) | Average Response Time (ms) |
---|---|
444 | 21 |
431 | 14 |
447 | 15 |
First result set
Next, we ran the same benchmark test using the API Management endpoint so requests were being proxied through API Management to the backend application. This scenario is an "end-to-end" or "system" test that is representative of how our API would be deployed in production. The results help us measure any latency or change in performance added by API Management and the Azure network. As we can see, the results are similar. This indicates that the net effect of API Management on the system performance at this design load is zero or very close to zero.
Throughput (RPS) | Average Response Time (ms) |
---|---|
443 | 15 |
441 | 14 |
436 | 10 |
Second result set
Finally, we ran a benchmark on a new "release" of our backend application. The new version of the API now returns a larger 1,500 byte payload, and we can see from the results that response times have increased significantly.
Throughput (RPS) | Average Response Time (ms) |
---|---|
361 | 600 |
370 | 518 |
367 | 585 |
Third result set
Assuming these results don't meet our performance objectives, we now know that remediation steps will need to be taken before the new release of our API can be deployed to production. For example, we might consider adding output caching, or scaling the App Service or API Management service, or look for ways to optimize the payload returned from the application code. In any case, we now have the tools to test any remediation approach (using the same structured, quantitative approach above) so that we can be sure that the new API version meets its performance objective before it's released.
Related resources to explore
- Performance tuning a distributed application
- Autoscaling
- Automate an existing load test with CI/CD
- Add caching to improve performance in Azure API Management
- Troubleshooting client response timeouts and errors with API Management
Continue reading...