Ed Glas's blog on VSTS load testing : Visual Studio 2010

New Web and Load Test Features in Beta2

edglas — Wed, 21 Oct 2009 20:42:00 GMT

In beta2 we added a number of new minor features, and a couple of major ones.

We also changed our licensing model, which you can read about here.

Here is a list of the minor features we added, along with a brief description of each. The last one in the list, compare two runs in Excel, is probably my most favorite feature in the entire release.

Area

Title

Description

Recorder

Recorder options in tools options

Starting URL supplies a default url to start recording from.

Two new options to control whether or not the recorder does hidden field matching, and whether or not it runs the dynamic parameter detection tool after recording completes.

Recorder

Recorder records all requests by default, fold dependents under page

In 2008, static dependents not found by the dependents parser were not recorded by default. Now they are.

Web test

search/replace

Search and replace in web test requests!

Playback

Add request bytes column

Request byte count is now shown in request list pane in playback (in addition to response byte count)

Playback

PreRequestDataBinding event

This new event fires before data binding occurs, allowing a plugin to dynamically change data bindings in a request.

Data binding

64 bit csv

This fixes a limitation we had in beta1, and now enables csv data sources on 64 bit machines.

Data binding

New APIs on WebTest to control data cursor position and reload data

Give full control over the cursor, enabling tight control of which user is using what data. Also reload enables refreshing a datasource during the test.

Load Test

Scrub counter sets

We refreshed the set of counters being collected, thresholds, and default counters.

Analyzer

Add standard deviation and 99^th percentile stats to load test analyzer

Analyzer

Graph instance 0 for CPU on local runs by default

Fixes the problem documented in this post.

Analyzer

Specify range for a counter (CPU range is 100)

Now you can specify a range for a given counter. Also percentage counters get a range of 100 by default.

Analyzer

Fixed range on perf graph

The performance graph now gets a default range set for it (range is rounded up response time of the slowest measurement)

Analyzer

Computer tags

Computer tags enable you to classify computers by function (e.g. web server, db server, etc.). Also enables run to run comparison even if the runs are on different physical machines.

Reporting

Compare two runs in Excel

This is a new, rich reporting option that enables you to compare performance and resource utilization data in two different runs.

Ed.

Getting Started with Visual Studio 2010 Beta 2

edglas — Tue, 20 Oct 2009 18:59:35 GMT

As you are probably already aware, Visual Studio Beta 2 is now available for download, see the landing page here.

For load testing, you will need to download and install:

Visual Studio Ultimate Beta 2 (x86)

Visual Studio 2010 Team Agents Beta 2 (x86)

Ultimate is the new equivalent of Team Suite, and contains Web performance tests and load tests.

Team Agents contains the controller and agents. Note that the x86 agent contains both 32 and 64 bit assemblies, so if you are looking for the 64 bit version of the load agent, this is the download you need.

As I posted here, we have moved to a virtual user license model. Ultimate is limited to 250 virtual users and one core of the CPU. To generate more than 250 users and use all cores on your machine, you will need to install a virtual user pack.

The most straight forward way to do this is to install a controller and agent (or multiple agents) on the same or separate machines.

We have re-architected the controller and agent setups such that setup does little more than lay down the assemblies, while a post-setup configuration tool does the heavy lifting to get the machine configured. The configuration steps are the ones that typically fail, so rather than having to reinstall the entire product and go back through all the configuration steps, if one part fails you only need to reconfigure that part. We have also enabled command line versions of both setup and the configuration tools.

Here is the controller config tool with options called out:

So you will need to fill in the controller service user using a user with counter collection access to your systems under test (same as in VS 2008), and then configure the load test database in the “Configure for load testing” section.

To add virtual users, start a Visual Studio Command prompt as Administrator (on Vista or Win7, just type Visual Studio Command Prompt in the search window, on XP you’ll find this under the Visual Studio group in the start menu). Then run this command line:

TestControllerConfig licenses /addkey:V10000

This will license the controller for 10,000 virtual users. Note that this license is temporary, and will only work in beta2. For RTM, we will only offer the 1,000 user license pack. If you have a Volume License, you will be able to get your license key from the volume licensing site, and then enter a count of licenses you want to install.

Once you have a virtual user license installed, that license can also be used for local runs (that is, the runs do not have to go through the controller).

If you just want to do local runs, and not use the controller and agent, you can run this from the command line to install the virtual user license:

VSTestConfig licenses /addkey:V10000

This enables you to exceed 250 users for local runs, and unlocks all the cores on your CPU.

Ed.

Introducing the Microsoft Visual Studio Load Test Virtual User Pack 2010

edglas — Tue, 20 Oct 2009 14:52:00 GMT

In Visual Studio 2010, we are changing our load test licensing model. In VS 2005 and VS 2008, we shipped the load agent, which was licensed by CPU on the machine generating load. In Visual Studio 2010 we are changing our license model to a virtual user based model. This new license model offers a number of advantages for you, which is ultimately why we chose to make the change:

Still way less expensive than our competition.
Enables better purchasing planning and allows an apples to apples price comparison with competitors.
Enables more flexible machine allocation for load generation.
Enables you to use agents for more than just load testing.

Still Way Less Expensive than our Competitors

Your initial reaction may be “Groan – that means that Microsoft is going to be just as expensive as those other tools!” But that is not the case at all, we were careful to maintain the per-virtual-user price we set out initially with VS 2005. With VS 2005 and VS 2008 load agents, the price was around $5,000 per load agent CPU, and our general guidance was that for a typical load test expect around 1,000 virtual users per CPU. We are not yet disclosing our pricing yet, but I can tell you our goal is to stay in this same ball park with VS 2010.

Better Planning and Comparison

With the VS 2008 load agent, we often get questions from customers doing purchasing planning asking “How many load agents do we need to purchase? We are planning to test up to 5,000 virtual users.” To which our answer with VS 2008 was “we do not know, it depends on your application and test scripts.” Of course this it difficult to know how many agents to purchase until actually doing testing, which was often too late. With virtual user licenses, purchasers will know exactly how many licenses to buy.

The old CPU-based license model also makes it very difficult to compare prices with competitive solutions. Most competitive products are licensed by virtual user and data collectors. If you know how many virtual users you want to test to, but do not know how many agents you will need to generate that number of users, it is impossible to make a comparison. With the Visual Studio Load Test Virtual User Packs 2010, if you know the virtual user count you want to test to, you will know exactly how many licenses to purchase, and be able to compare directly with competitive solutions you may consider.

Enables More Flexibility with Load Generators

The virtual user license offers flexibility with the machines you use as load generators.

With VS 2008, to get the full value of the per-CPU license, you had to run on fast CPUs, often requiring new hardware purchases. Also the license was locked to a particular machine, restricting which machines you could use as load generators. With Visual Studio 2010, you can now run virtual users on older hardware with no “penalty”, as you can line up as many older machines as you wish to generate the target number of virtual users. Also, the virtual user license is now tied to the controller, not the agent machine, so you are allowed to move agent machines in and out of the agent pool.

For some tests, you may run out of some other resources before running out of CPU. The most common case for this in VS 2008 was load agents running out of memory before CPU (especially for web tests with very large post bodies or very large responses), which prevented customers from realizing the full value of their license. With the virtual user license, if this happens you can simply add more machines to the agent pool to achieve the load you are targeting.

Visual Studio 2010 Agents: More than just Load Generators

With the high cost of the agent license, we did not have a viable solution for distributed functional testing. In Visual Studio 2010, agents are now included at no extra cost with VS Premium, Ultimate, and Test Elements. We’ve added Data Collection facility to agents, and added new UI automation test type, which opens up a bunch of new scenarios for using agents.

What if I already have VS 2008 Load Agents?

If you have purchased VS 2008 load agents with software assurance, you will be able to migrate your existing licenses to a commiserate set of Virtual User License Packs. We will iron out the details and mechanics of this once we get closer to RTM.

Conclusion

You can see there are many advantages to you with the new licensing model.

Ed.

Using The Virtual User Activity Chart to Understand the VS Load Engine

edglas — Mon, 19 Oct 2009 03:07:14 GMT

In dev10 we’ve added a new Virtual User Activity Chart that shows virtual user activity. Each row on the chart is a virtual user.

Covering the Basics of the Chart

See Sean Lumley’s post VSTS 2010 Feature- Load test virtual user activity visualization to see an over view of the Virtual User Activity Chart.

Below is the test view of the chart, which shows which shows which test a user is running. This load test has 25 user constant load, with three unit tests in it: Fast (takes 1 second to execute), Medium (takes 5 seconds to execute), and Slow (takes 10 seconds to execute), using the Sequential Test Mix (which I will go into more detail on later in the post).

Zooming in and using Paint to group 5 users at a time for easy counting, the chart shows 25 rows, one for each user.

Zooming in using the zoom bar on the chart, you can see user 0 first ran the fast test, then the medium test, then the slow test:

Which is confirmed by hovering over the purple Medium test bar:

This view shows you exactly what each virtual user was executing during the test, so it will enable you to see patterns of user activity, load patterns, and correlate failed or slow tests and requests with other virtual user activity.

Another use of this view is it really helps you understand what the load engine does.

In this post, I’ll review various properties in the load engine and show the details view so you can see the effect these properties have.

Understanding Test Mixes

First, I’ll show different test mixes and their corresponding details view, using tests that will accentuate the differences in the mixes.

I’ll use three tests, long, medium and short. Long runs for 10 seconds, medium for 5 seconds, and short for 1 second.

Test Mix Based on Number of Tests Started

Here is the Activity Chart for this test mix, with 34% Fast, 33% Medium, 33% Slow. With this mix, as each user finishes a test, the user randomly selects the next test based on the test weightings.

One think you’ll notice is the predominance of Yellow, and lack of Blue. You wouldn’t think it to look at the user details view, but at the end of this test, the count for each test type is roughly the same. Since the Slow test runs 10x slower than the fast test, there is 10x more yellow than blue.

Name

Scenario

Total Tests

Avg. Test Time (sec)

Slow

Scenario1

912

10.0

Medium

Scenario1

934

5.00

Fast

Scenario1

1,012

1.00

34/33/33 Test Mix Based on the Number of Virtual Users Starting Tests

What this mix is doing is trying to keep, and any given time, the mix of virtual users running the tests to hit the percentage specified. Given the fast test starts and ends so quickly, you would expect a lot of more iterations of the fast test which is exactly what we see, a lot more blue:

Name

Scenario

Total Tests

Avg. Test Time (sec)

Slow

Scenario1

472

10.0

Medium

Scenario1

952

5.00

Fast

Scenario1

5,329

1.00

Test Mix Based on User Pace

In this test, I’ll run the same three tests, and specify they each run 120 times per hour. I chose 120 since it will take 16 seconds to run all three tests, so if I run each twice a minute, I should get a nice graph with about half the time spent running tests. Of course, this mix is not intended to be used this way, it should be used to model your user behavior (how many times an hour do you expect a user to do a particular action).

Now we see a very different graph, with a lot of pacing time between iterations:

Name	Scenario	Total Tests	Avg. Test Time (sec)
Slow	Scenario1	174	10.0
Medium	Scenario1	177	5.00
Fast	Scenario1	183	1.00

Test Mix Based on Sequential Test Order

With this mix, each user simply runs the tests in the order you specify in the mix, which is easy to conceptualize and super-easy to see:

The Effect of Think Time Between Iterations and Warmup Duration Properties

As you can see from this last detail view, this test has all users going lock step through the application, all doing exactly the same thing at the same time. While this may be useful for doing stress testing, it is not realistic user behavior, which is why we have the other test mixes which do a better job of varying what users are doing.

Another way to vary what users are doing and introduce pacing into a scenario is by using Warmup Duration and Think Time Between Iterations. Think Time Between Iterations is especially important if your test only has a single web test in it (thus no think time), or you are running a unit test with no pacing (for unit tests this is implemented via sleep calls, but never use sleep calls in a web test).

Here’s what happens when I put a 30 second warmup in the last test:

Now you can see the virtual users stagger their start over 30 seconds.

And here is the same test (no warmup) with a 5 second think time between iterations:

The variations are due to the Normal Distribution think time profile that is applied by default. This profile will vary the specified think time using a normal distribution (bell curve).

The Effect of Percentage of New Users

Now I am going to clear the delay between iterations, and use 100% new users. What I’d expect here is that each user will only run one test, leading to a lot of rows in the view, which is exactly what happens:

Scrolling up vertically one screen shows this:

So you can see the load engine repeatedly starting new users. As one group of 10 users completes, the next group of 10 starts. New users have a clean cache and cookie container, which matter for web tests.

Different mixes interact differently with the percentage of new users:

Sequential Mix as you can plainly see above, we made the design decision on sequential mix to have each new user run each test once.
With Test Mix Based on Sequential Test Order and Test Mix Based on Number of Tests Started, each user runs one test.
Test Mix Based on User Pace ignores the percentage of new users altogether, as the mix relies on a repeat user to establish the proper pacing.

The Effect of Test Mix Init and Terminate Tests

Init and terminate tests one once per each user. Switching my test back to 0 percent new users, and adding an init and terminate test shows the following:

Where the legend shows

This clearly shows the init test running once per virtual user, as the very first test executed, exactly as it should.

But wait! Where is the terminate test? It is not in the legend, and does not show up scrolling the view all the way to the end of the test. This shows the terminate test only gets called when a cooldown period is configured in the load test. Re-running with a 30 second cool down demonstrates this:

Understanding Load Patterns with the Virtual Users Activity Chart

Each of the above tests uses a constant user load of 25 users. Now I’ll use the chart to show how Step Load and Goal-based load based loads add users. For these tests, I’m going to use just one test in the mix, the 1 second fast test, to draw attention to the load pattern rather than the test mix.

Step Load

Here’s the Activity Chart for a step load pattern adding 10 users every 10 seconds until 100 users is reached. Notice the vertical scroll bar to scroll up to see the remaining users. Zooming in on the first 30 seconds shows the first 30 users created.

The step load Step Ramp Time property enables you to ramp the users added at each step. I highly recommend using this for every step load pattern, particularly if you are using large steps. Here I set a step ramp of 5 seconds, which shows users ramped up for each step over 5 seconds.

Goal Based Load

For a goal based load pattern, it is important that the counter you are seeking the goal on has a direct correlation with the load at each interval. So for this demonstration I put a tight loop in my test to provide a direct correlation between CPU on the load agent and the virtual user count. For this test, I also did a step load before hand to see what a good CPU range would be.

Here’s user load:

Let’s look at the Activity Chart for the areas where the dips are:

Here you can see virtual users stopping as they finish an iteration in order to achieve a lower load.

Scrolling up, you can see that new users are added to the test to move the load back up.

One thing that surprised me here is that users are not recycled. That looks like a bug to me!

Conclusion

Hopefully this post unravels some of the properties of the load engine, and it shows you just one of the ways you can leverage the new Virtual User Activity Chart in dev10 to understand exactly what your load test is doing!

Read this Before Running a Load Test with Dev10 Beta 1

edglas — Sun, 24 May 2009 08:06:00 GMT

If you install the dev10 beta 1 on the same machine as VS 2008, and you are running load tests, read this.

When you run the first load test, the load test will check the schema of the load test database, and if it is the 2008 schema it will update the schema in place.

This will create problems for you in VS 2008: you will no longer be able to import and export results.

The workaround is to use two separate databases for your results, one for VS 2008, one for VS 2010. The new Excel reporting feature in dev10 looks for the load test database (we’re fixing that in beta2), so the best workaround is to rename the VS 2008 load test database.

Before running any 2010 load test on your machine, rename the VS 2008 load test database:

Close all Visual Studio instances.
Run a Visual Studio command prompt as Administrator (can be 2008 or 2010).
Run
sqlcmd –S .\sqlexpress
1> EXEC sp_renamedb 'LoadTest', 'LoadTest2008'
2> GO
Start Visual Studio 2008
Run Test | Administer Test Controllers
Change the load test results store to the renamed database
1. Click …
2. Under “Select or Enter a Database Name” type “LoadTest2008”

If you run VS 2010 tests prior to renaming the database, you will need to remove the dev10 results from the database:

From Visual Studio 2010, open a load test.
Click Open and Manage Test Results icon
Select the 2010 test results and click export.
Delete the 2010 test results.
Go through the steps above to create two databases.
Open VS 2010 and use Open and Manage Test Results to import the 2010 results back into the database.
Delete the extra columns created in your database. We’ll blog a script for doing that soon.

Note the same thing will happen during setup if you install the Test Controller – controller setup will update the existing 2008 load test database. So be sure to rename the database before running controller setup.

Ed.

dynaTrace Leverages new Extensibility Points to Provide Deep Integration with Dev10

edglas — Wed, 20 May 2009 20:25:30 GMT

A fantastic example of a partner taking advantage of the new extensibility points in Web test playback and load tests in dev10 is provided by dynaTrace. dynaTrace has developed an extension that provides enables you to drill from a particular request to the PurePath for that request. The PurePath provides a tiered performance breakdown for the request, detailing how much time the request spent at each tier of your application. Check out Andreas’s post on the integration (including some code samples!).

Ed.

Dev10 Beta 1 Available!

edglas — Mon, 18 May 2009 18:15:00 GMT

Beta 1 was released to the web today for MSDN subscribers. If you are a MSDN subscriber, you can download the Beta today from here.

We have been very busy on this release of late (one reason why my blog posts have slowed down lately :))

My team and I will be blogging about all the great new features in dev10 beta 1 over the coming weeks.

As in the past, I will compose a list of links to any posts on the new features and update this blog post with the links as we get them.

New Web Test Features in Dev10:

These features are about making it easier to get web test playback working:

Extensible recorder plugins for modifying recorded web tests (including adding custom dynamic parameter correlation)
Find in playback
View recording log in playback
Jump to request from playback
Add extraction rule from playback
WebTestResult in a separate file for better scalability
New API for web test result
New extensibility API in playback enables UI extensions to playback

File Upload record/playback just works

Uploaded file added to project and deployed
Option for generating a unique file name on each iteration

Looping and branching in web test
New dialog for editing think times, goals, reporting string across all requests
New validation and extraction rules for inner text and select tags
Test level validation rule for pages meeting response time goal

New features in Load Test in Dev10:

Logging and data collection improvements

Log entire test result on test failure or log entire result every n requests
Remote profiler integration
Event Log Data Collection
Extensible data collectors to support more platforms

Analysis Improvements

Load test virtual user activity visualization, also here.

Reporting

Run to run perf reports in Microsoft Excel

Performance testing

Run automated UI tests as performance tests
True network emulation

More flexible load modeling

New Sequential Test Mix
Delay start on scenario
Iteration count on scenario
Select which agent(s) to run the scenario on
Disable during warmup

Multiple load test plugins
64 bit test execution

In addition to continuing our testing and bug fixing efforts, right now we’re working on an additional set of features for beta 2. I’ll have to sit on those until beta2. :)

Ed.