Category Archives: Performance

XLT 4.1.8 Update Release

XLT 4.1.8 has been released. It contains a couple of improvements and bug fixes. You can get the latest version here: https://lab.xceptance.de/releases/xlt/4.1.8/.

These are the two most important changes.

Optimized scanning CSS rules

XLT simulates the download behavior of a browser as close as possible. With com.xceptance.xlt.css.download.images set to onDemand, XLT carefully considers the download of image resources referred to in CSS files. It checks all CSS rules if they apply to all elements in the DOM tree and, if so, extracts the resource information for later download. This process was previously very expensive and ran for up to several seconds when a page was complex and a lot of CSS rules had to be matched. This has been optimized heavily.

Configurable retry behavior for keep-alive connections

If persistent connections (keep-alive) are enabled, test cases might failed sporadically with an exception such as:
java.lang.RuntimeException: org.apache.http.NoHttpResponseException: The target server failed to respond

This exception occurs only when the request was sent while the connection was about to be closed on the server side at the same time.

The underlying HttpClient retries requests if the connection was closed by the server. However, by default it retries only idempotent operations such as GET and HEAD, but not POST and PUT. In order to avoid these connection errors, the user can now configure whether non-idempotent operations are to be retried as well. Common browsers seem to use the same behavior and most certainly assume that the server did not start to process the request when it closes the connection immediately without responding with a proper HTTP status code.

A new boolean property com.xceptance.xlt.http.retry.nonIdempotentRequests in config/default.properties controls whether or not operations such as PUT and POST are retried in case of certain networking problems. Note that idempotent operations are always retried.

Get the right load mix out of a few numbers

When testing ecommerce applications on SaaS environments, you often do not get enough numbers from clients because they simply do not know these numbers or only a few. One reason for that is, that the client simply have not had any only presence before. Often the client also does not have detailed numbers, because the previous hoster or the IT department just holds them back or simply cannot get to these numbers.

So what to do, when you do not know every detail about the current or future load pattern? We are describing one approach below that was very successful so far and always yielded satisfying results.

What we need

  • Visits per peak hour (example 10k)
  • Page views per peak hour (example 100k)
  • Orders per peak hour (example 200 orders)
  • Optionally we can use the conversion rate to get from visits to orders or vice versa.
  • Optionally we can take searches, “add to cart” operations, user registrations, and so on into account.

The mentioned scenarios are typical ecommerce scenarios and look like that. We will not talk about smaller scenarios such as address editing for a registered user.

  • TSingleClickVisit: Enters the store only, does not move beyond the start page
  • TBrowsing: TVisitor plus category and product browsing
  • TSearch: TVisitor plus keyword search plus browsing of the result
  • TAdd2Cart: TBrowsing plus add to cart operations
  • TGuestCheckout: TAdd2Cart plus checkout without an order placement (anonymous user)
  • TGuestOrder: TAdd2Cart plus full checkout (anonymous user)
  • TRegisteredCheckout: TAdd2Cart plus checkout without an order placement (registered customer)
  • TRegisteredOrder: TAdd2Cart plus full checkout (registered customer)
  • TRegistration: Account creation

What we assume

Ecommerce sites follow similar patterns and with a few exceptions, such as special promotions, certain behavioral patterns are nearly identical. So for instance, about 50% of all checkouts are stopped before the order is placed. About 20 to 50% of all created carts aren’t checked out at all.

What we calculate

Based on these assumptions, we put together a fairly simple but sufficiently accurate load mix. Of course, we can also analyze the current log files and try to come up with something more precise, but that will be a snapshot only. Traffic is very volatile and so we should be very generous when setting up this mix.

Since we do not take any daily averages as base but the peaks, we will have a pretty comfortable buffer for our daily ecommerce life anyway.

Bottom-Up

Let’s say, 200 orders are set as goal. Splitting them 50/50  between registered and anonymous users, we get 100 visits of each type. All numbers are per hour of course.

  • TGuestOrder = 100
  • TRegisteredOrder = 100

As a next step, we take our 50% checkout abandonment rate into account. We have 200 checkouts per hour that are stopped and 200 that run through and turn up as orders (as counted previously). So we need to add 200 visits. And because these visitors can either run with their preset account or without, we split them up in 100 guest and 100 registered checkout attempts.

  • TGuestCheckout = 100
  • TRegisteredCheckout = 100
  • TGuestOrder = 100
  • TRegisteredOrder = 100

This gives us 400 visits per hour that go into the checkout. We now assume a low cart to checkout conversion rate, about 20% for instance, and so we take 400 checkout visits * 5 and get 2,000 visits that involve cart usage. Since we already have 20% converted into checkouts, we have 2,000 minus 400 visits that use the cart.

  • TAdd2Cart = 1,600
  • TGuestCheckout = 100
  • TRegisteredCheckout = 100
  • TGuestOrder = 100
  • TRegisteredOrder = 100

We also know that many users do not continue after hitting the home page or any landing page. Let’s add some of these users now.

  • TSingleClickVisitor = 1,000
  • TAdd2Cart = 1,600
  • TGuestCheckout = 100
  • TRegisteredCheckout = 100
  • TGuestOrder = 100
  • TRegisteredOrder = 100

But wait, what are we missing? Well, we have not registered any new accounts yet. Didn’t we? We did, because the registered checkout creates accounts if required and reuses them several times. But to get a more substantial customer growth, we simply add 200 visits that run registrations:

  • TRegistration = 200
  • TSingleClickVisitor = 1,000
  • TAdd2Cart = 1,600
  • TGuestCheckout = 100
  • TRegisteredCheckout = 100
  • TGuestOrder = 100
  • TRegisteredOrder = 100

What is left to do? Well, we do not have any “I am just looking around”-visitors yet. We know that our total visit count is 10,000 and we already assigned 3,200 of these to cart, checkout, and registration, so we have 6,800 visits left we can now use for something else. Depending on the shop type (large store, small store etc), people tend to use search more or less. To put enough stress on search and refinements, we simply assume 50% of all people like to search. Thus the missing 6,800 visits will be 3,400 catalog browser visits and 3,400 visits with usage of search before browsing the search result.

The total mix is:

  • TBrowsing = 3,400
  • TSearch = 3,400
  • TRegistration = 200
  • TSingleClickVisitor = 1000
  • TAdd2Cart = 1,600
  • TGuestCheckout = 100
  • TRegisteredCheckout = 100
  • TGuestOrder = 100
  • TRegisteredOrder = 100

Wait… where are my concurrent users? This is simple: “concurrent users” is an inaccurate way of describing traffic, so we have not used that number yet. Why is that?

To get to the bottom of that, we simply check how long a visit takes. Depending on the shop, an average visit might take 2 to 4 minutes. Successfully shopping might take 15 minutes. If we expect about 10 page views per visit and a page view takes 1 second to load and 20 seconds to read it (already a really really high number for an average), a visit would take 10 * 1 second + 9 * 20 seconds = 190 seconds.

Let’s go with the 190 seconds for a visit on average. If we just could serve one visitor at a time, we could serve 60 minutes (3600 seconds) / 190 seconds per visits = 19 visitors per hour. But because we would like to serve 10,000 per hour, we have to deal with 10,000 / 19 = 526 visitors at the same time. This is the famous concurrent user number.

If we now double the think time, we have 1,052 concurrent users/visitors. If we cut it down to 1 second think time, we will get a visit length of 19 seconds and therefore 10,000 visits / (3600 seconds / 19) = 53 concurrent visitors.

So we already have three different “concurrent user” numbers and are still simulating the same traffic. This shows that the number of concurrent users is a pretty questionable way of describing traffic.

It does not matter which number we take, because most of the time the servers will see the same traffic. Because we run against a SaaS environment that serves a multiple of other customers at the same time and is sized to serve the peak traffic for all customers at the same time, we have plenty of comfortable room around us. This permits us to run with 53 concurrent visitors for most of the testing. This will save us client hardware resources for the load generation. e.g. saves us money. We are basically only interested in the runtime of requests and not if the environment can handle that, because it can.

The goal of this test is to demonstrate that the implementation on the SaaS platform is efficient, not that the SaaS platform itself is fast and stable, because this is guaranteed by design and contract. Testing this would require way more traffic and generate huge costs, because the environment would suddenly no longer be a shared one but exclusively used for this testing purpose.

When finalizing the entire test and all tests turned out good, we are going to turn up the concurrent user count to 530 users and compare the result with the previous measurements. Just to satisfy the traditional test expectations.

Does that work for you?

Hope that gives you an idea how to come up with a nice user mix for testing without having too much data in the first place. Comments welcome.

Nice reading: CSS3 vs. CSS

Nice article about the advantages of CSS3 in terms of coding as well as download speed: CSS3 vs. CSS: A Speed Benchmark.

I believe in the power, speed and “update-ability” of CSS3. Not having to load background images as structural enhancements (such as PNGs for rounded corners and gradients) can save time in production (i.e. billable hours) and loading (i.e. page speed). At our company, we’ve happily been using CSS3 on client websites for over a year now, and I find that implementing many of these properties right now is the most sensible way to build websites.

Until today, all of that was based on an assumption: that I can produce a pixel-perfect Web page with CSS3 quicker than I can with older image-based CSS methods, and that the CSS3 page will load faster, with a smaller overall file size and fewer HTTP requests. As a single use case experiment, I decided to design and code a Web page and add visual enhancements twice: once with CSS3, and a second time using background images sliced directly from the PSD. I timed myself each round that I added the enhancements, and when finished, I used Pingdom to measure the loading times.

More…

Enjoy reading.

Load Testing Web Applications – Do it on the DOM Level!

This article was first published in the June 2010 issue of the magazine Testing Experience.

HTTP-level tools record HTTP requests on the HTTP protocol level. They usually provide functionality for basic parsing of HTTP responses and building of new HTTP requests. Such tools may also offer parsing and extraction of HTML forms for easier access to form parameters, automatic replacement of session IDs by placeholders, automatic cookie handling, functions to parse and construct URLs, automatic image URL detection and image loading, and so on. Extraction of data and validation are often done with the help of string operations and regular expressions, operating on plain HTTP response data. Even if HTTP-level tools address many load testing needs, writing load test scripts using them can be difficult.

Challenges with HTTP-Level Scripting

Challenge 1: Continual Application Changes

Many of you probably know this situation: A load test needs to be prepared and executed during the final stage of software development. There is a certain time pressure, because the go-live date of the application is in the near future. Unfortunately, there are still some ongoing changes in the application, because software development and web page design are not completed yet.

Your only chance is to start scripting soon, but you find yourself struggling to keep up with application changes and adjusting the scripts. Some typical cases are described below.

  • The protocol changes for a subset of URLs, for example, from HTTP to HTTPS. This could happen because a server certificate becomes available and the registration and checkout process of a web shop, as well as the corresponding static image URLs, are switched to HTTPS.
  • The URL path changes due to renamed or added path components.
  • The URL query string changes by renaming, adding or removing URL parameters.
  • The host name changes for a subset of URLs. For example, additional host names may be introduced for a new group of servers that delivers static images or for the separation of content management URLs and application server URLs that deliver dynamically generated pages.
  • HTML form parameter names or values are changed or form parameters are added or removed.
  • Frames are introduced or the frame structure is changed.
  • JavaScript code is changed, which leads to new or different HTTP requests, to different AJAX calls, or to a new DOM (Document Object Model) structure.

In most of these cases, HTTP-level load test scripts need to be adjusted. There is even a high risk that testers do not notice certain application changes, and although the scripts do not report any errors, they do not behave like the real application. This may have side effects that are hard to track down.

Challenge 2: Dynamic Data

Even if the application under test is stable and does not undergo further changes, there can be serious scripting challenges due to dynamic form data. This means that form field names and values can change with each request. One motivation to use such mechanisms is to prevent the web browser from recalling filled-in form values when the same form is loaded again. Instead of “creditcard_number”, for example, the form field might have a generated name like “cc_number_9827387189479843”, where the numeric part is changed every time the page is requested. Modern web applications also use dynamic form fields for protection against cross-site scripting attacks or to carry security-related tokens.

Another problem can be data that is dynamically changing, because it is maintained and updated as part of the daily business. If, for example, the application under test is an online store that uses search-engine-friendly URLs containing catalog and product names, these URLs can change quite often. Even worse, sometimes the URLs contain search-friendly catalog and product names, while embedded HTML form fields use internal IDs, so that there is no longer an obvious relation between them.

Session IDs in URLs or in form fields may also need special handling in HTTP-level scripts. The use of placeholders for session IDs is well supported by most load test tools. However, special script code might be needed, if the application not only passes these IDs in an unmodified form, but also uses client-side operations on them or inserts them into other form values.

To handle the above-mentioned cases, HTTP-level scripts need manually coded, and thus unfortunately also error-prone, logic.

Challenge 3: Modeling Client-Side Activity

In modern web applications, JavaScript is often used to assemble URLs, to process data, or to trigger requests. The resulting requests may also be recorded by HTTP-level tools, but if their URLs or form data change dynamically, the logic that builds them needs to be reproduced in the test scripts.

Besides this, it can be necessary to model periodic AJAX calls, for example to automatically refresh the content of a ticker that shows the latest news or stock quotes. For a realistic load simulation, this also needs to be simulated by the load test scripts.

Challenge 4: Client-Side Web Browser Behavior

For correct and realistic load simulations, the load test tool needs to implement further web browser features. Here are a few examples:

  • Caching
  • CSS handling
  • HTTP redirect handling
  • Parallel and configurable image loading
  • Cookie handling

Many of these features are supported by load test tools, even if the tools act on the HTTP level, but not necessarily all of them are supported adequately. If, for example, the simulated think time between requests of a certain test case is varied, a low-level test script might always load the cacheable content in the same way – either it was recorded with an empty cache and the requests are fired, or the requests were not recorded and will never be issued.

DOM-Level Scripting

What is the difference between HTTP-level scripting tools and DOM-level scripting tools? The basic distinction between the levels is the degree to which the client application is simulated during the load test. This also affects the following characteristics:

  • Representation of data: DOM-level tools use a DOM tree instead of simple data structures.
  • Scripting API: The scripting API of DOM-level tools works on DOM elements instead of strings.
  • Amount and quality of recorded or hard-coded data: There is much less URL and form data stored with the scripts. Most of this data is handled dynamically.

DOM-level tools add another layer of functionality on top. Besides the handling of HTTP, these tools also parse the contained HTML and CSS responses to build a DOM tree from this information, similar to a real web browser. The higher-level API enables the script creator to access elements in the DOM tree using XPath expressions, or to perform actions or validations on certain DOM elements. Some tools even incorporate a JavaScript engine that is able to execute JavaScript code during the load test.

Advantages

DOM-level scripting has a number of advantages:

  • Load test scripts become much more stable against changes in the web application. Instead of storing hard-coded URLs or data, they operate dynamically on DOM elements like “the first URL below the element xyz” or “hit the button with id=xyz”. This is especially important as long as application development is still ongoing. As a consequence, you can start scripting earlier.
  • Scripting is easier and faster, in particular if advanced script functionality is desired.
  • Validation of result pages is also easier on the DOM level compared to low-level mechanisms like regular expressions. For example, checking a certain HTML structure or the content of an element, like “the third element in the list below the second H2” can be easily achieved by using an appropriate XPath to address the desired element.
  • Application changes like changed form parameter names normally do not break the scripts, if the form parameters are not touched by the script. But, if such a change does break the script because the script uses the parameter explicitly, the error is immediately visible since accessing the DOM tree element will fail. The same is true for almost all kinds of application changes described above. Results are more reliable, because there are fewer hidden differences between the scripts and the real application.
  • CSS is applied. Assume there is a CSS change such that a formerly visible UI element that can submit a URL becomes invisible now. A low-level script would not notice this change. It would still fire the old request and might also get a valid response from the server, in which case the mismatch between the script and the changed application could easily remain unnoticed. In contrast, a DOM-level script that tries to use this UI element would run into an error that is immediately visible to the tester.
  • If the tool supports it, JavaScript can be executed. This avoids complicated and error-prone re-modeling of JavaScript behavior in the test scripts. JavaScript support has become more and more important in recent years with the evolution of Web 2.0/AJAX applications.

Disadvantages

There is one disadvantage of DOM-level scripting. The additional functionality needs more CPU and main memory, for instance to create and handle the DOM tree. Resource usage increases even more if JavaScript support is activated.

Detailed numbers vary considerably with the specific application and structure of the load test scripts. Therefore, the following numbers should be treated with caution. Table 1 shows a rough comparison, derived from different load testing projects for large-scale web applications. The simulated client think times between a received response and the next request were relatively short. Otherwise, significantly more users might have been simulated per CPU.

Scripting Level Virtual Users per CPU
HTTP Level 100..200
DOM Level 10..30
DOM Level + JavaScript execution 2..10

If you evaluate these numbers, please keep in mind that machines are becoming ever more powerful and that there are many flexible and easy-to-use on-demand cloud computing services today, so that resource usage should not prevent DOM-level scripting.

Conclusion

Avoid hard-coded or recorded URLs, parameter names and parameter values as far as possible. Handle everything dynamically. This is what we have learned. One good solution to achieve this is to do your scripting on the DOM level, not on the HTTP level. If working on the DOM level and/or JavaScript execution are not possible for some reason, you always have to make compromises and accept a number of disadvantages.

We have created and executed web-application load tests for many years now, in a considerable number of different projects. Since 2005, we have mainly used our own tool, Xceptance LoadTest (XLT), which is capable of working on different levels and supports fine-grained control over options like JavaScript execution. In our experience, the advantages of working on the DOM level, in many cases even with JavaScript execution enabled, generally by far outweigh the disadvantages. Working on the DOM level makes scripting much easier and faster, the scripts handle many of the dynamically changing data automatically, and the scripts become much more stable against typical changes in the web application.

Ronny Vogel is technical manager and co-founder of Xceptance. His main areas of expertise are test management, functional testing and load testing of web applications in the field of e-commerce and telecommunications. He holds a Masters degree (Dipl.-Inf.) in computer science from the Chemnitz University of Technology and has 16 years of experience in the field of software testing. Xceptance is a provider of consulting services and tools in the area of software quality assurance, with headquarters in Jena, Germany and a branch office in Cambridge, Massachusetts, USA.

XLT – Garbage Collector details visualized

Today we want to give you a small preview of an upcoming XLT feature. Most of you probably know that XLT features agent statistics. These statistics help you to keep an eye on the health of the test execution engines (agents) to ensure that you do not influence the test results by providing insufficient hardware or by applying no or incorrect settings.

Most modern programming languages are virtual machine based and these machines have knobs you can turn to adjust their behavior according to your requirements. XLT runs on Java and so all the things you might have already learnt from tuning your Java-based servers apply to XLT as well. If you do not have experience in tuning your Java-based servers, you will learn a lot that can be applied to your servers and help you to increase performance.
Continue reading XLT – Garbage Collector details visualized

Browser Cache Usage Study by Yahoo

Today, I rediscovered this nice article about browser cache usage: Performance Research, Part 2: Browser Cache Usage – Exposed!. It gives you a pretty good idea about the average cache usage. Bottom line: Optimize your site for no cache hits at all and you are good.

40-60% of Yahoo!’s users have an empty cache experience and ~20% of all page views are done with an empty cache. … It says that even if your assets are optimized for maximum caching, there are a significant number of users that will always have an empty cache. …reducing the number of HTTP requests has the biggest impact on reducing response time. The percentage of users with an empty cache for different web pages may vary, especially for pages with a high number of active (daily) users. However, we found in our study that regardless of usage patterns, the percentage of page views with an empty cache is always ~20%.

Burn-in test of XLT 4.0

If you write a performance testing software, your first obligation is a performance test for and with that software. So before we can ship XLT 4.0, we have to make sure that it can hold up to its promises. Test software has to be tested too, so to speak. Today we provisioned a bunch of Amazon-EC2 boxes with 30 test agents in total. These charts are from a short test to see where we can get to, when ignoring any target numbers. This was a short and violent test. Just hammer the system under test and see if it can and will recover.

HitsPerSecond

A short ramp up period in the beginning of the test and afterwards we kept a steady load factor. For the steady phase we reached more than 11,000 hits per second.

Concurrent Users

The target system was seeing about 2,200 concurrent users during the peak of the test.

Received Bytes per Second

During that test, the network was transporting about 95 Mbytes/s inbound traffic, this is a network utilization of about 760 Mbit/s. Amazing that the EC2 boxes in the EU data center can handle that traffic. We used just 10 boxes and each box has a 100Mbit network, so the overall limit must have been reached we think.

By the way, the target system recovered easily and was able to serve its normal duties without any problems. But this test clearly showed the limits in terms of throughput. But this test did not show any limits of XLT 4.0, because neither the load boxes in terms of CPU nor the reporting had any problems with this test size.

Reasons for a Test Environment

People asked for a rough guide for running load tests against their live site and whether this is a good idea at all.

Well, let me first say that test environments exist for exactly that purpose. So if you already have something live, of course you used it before going live for load testing, and now you cannot run another test there… yes, you need a test environment. This is well spent money for several reasons:

  • Created data will not pollute your installation aka log entries, analytics data, orders in the database, created users, and so on.
  • You do not have to take your live site down for testing.
  • Problems during testing will not leave your live site down or take it down.
  • Your hosting company or IT department might charge for all test traffic against the live site as it would have been real traffic (revenue share, bandwidth utilization), so having your own testing realm makes expenses more predictable.
  • You can easily change code and retest when changes are necessary. You can profile, you can debug.
  • You do not risk problems with traffic going to live systems, such as payment due to configuration or testing errors. If you already had items from test orders piling up in your office, you know what I mean.

So, get yourself a test environment. Spend the money and enjoy the freedom to measure, debug, and tune.

P.S. Of course, there are situations, where you cannot replicate a live environment or you cannot find application problems in a test setup due to totally different timing behavior… Well, this indicates only that your live environment is already screwed.

What is the state of the G1 Garbage Collector?

Because I do not know what is the current state of the Java G1 Garbage Collector, I decided to try G1 with JDK6u20. Somehow I was disappointed because after a short moment of predictable GC performance, the entire VM stopped and some major collection was running. You can easily see that in the charts of that run. Right around 20:09:45, the threads were stopped and the entire VM behaved ugly.

The G1 stop the world

So, the G1 is not yet ready for production, of course nobody stated that it is ready for production. If I read the release notes of JDK6u21 correctly, it delivers plenty of G1 changes, so I might try that soon.