You can't have it all – so the saying goes – and when it comes to application optimization, that is typically true. There are trade-offs that are made when deciding what optimization needs to be deployed. Three of the main aspects when optimizing an application to consider are:
- What performance impact an optimization will have. Is the optimization going to help the majority of customers, the majority of content, and will it cause a double-digit performance improvement?
- How reliable the optimization is. Will implementing a new technique cause an application to break? Does the optimization cause some pages to load faster and other pages to load slower, or have first-time visits improved but repeat visits become slower?
- How much time and effort is required for implementation. Will you have to spend time creating a whitelist or blacklist to ensure the site continues to function for all users? What happens when you change your site? Do you have to reconfigure all of the optimizations?
When it comes to optimization techniques, the list of what can be done is constantly growing, and not each technique has the same ROI.
The grid above attempts to chart some of the most common optimizations available through FEO- and server-based along three variables. The y-axis is the measurement of the performance impact, the x-axis is the measurement of the reliability of the implementation, and the size of the dot reflects the amount of time required for implementation (the larger the dot, the more time-consuming).
Low Performance, Low Reliability
These features have minimal performance impact and implementing them can result in adverse consequences. The goal of in-lining and concatenation of resources is to reduce the number of round trips required to load a page, as generally the more round trips required, the longer it takes to load a page. But these techniques only apply to first-time visitors, or those with an empty cache. The bigger downside of these techniques is that they can “break” caching for repeat visitors or a subsequent page views in a browsing session.
Figure 1: Prior to concatenation
Figure 2: After concatenation
Similarly, in-lining resources can also negatively impact caching. In-lining takes external files and embeds them in the HTML to eliminate the round trip. HTML is typically not cached by a browser and is retrieved for each request, resulting in repeat viewers getting worse performance after in-lining is implemented. Many people will choose to only in-line smaller files, but this adds layers of complexity to determining the appropriate threshold and the creation of the whitelist.
With increased adoption of HTTP/2, both of these techniques will be deprecated, as concurrency is a core feature of the HTTP/2 specification. Spending time to implement techniques that will not be needed in an HTTP/2 world does not seem worth the effort, especially given the low performance impact.
High Performance, Low Reliability
These optimizations have a higher performance impact but still have challenges in terms of reliability. While all of these optimizations attempt to improve the performance of images, they all go about it in different ways. Improving the performance of images often results in high performance gains, as they make up such a large percentage of page weight. The time required for implementation may be worth the performance gains. One thing to remember with each of these is the ongoing care and feeding that is required; any web site change means re-testing and potentially rewriting code to execute correctly.
Image sprites were one of the first optimizations on the scene. The goal was to reduce the number of round trips by combining all the images into a single file and use CSS pointers to tell the browser which part of the image to display. Having to create a new sprite and update the CSS when images change on your application reduces the reliability of images on your site changing frequently.
Prefetching images is anticipating what a user will click on next and populating the browser's cache with that content. As images are highly cacheable, they are frequently pre-fetched. Requesting images before a user needs them can speed up performance if the prediction was correct. Predicting human behavior is not easy; you may end up sending content to the user that wasn’t needed, wasting precious bandwidth. It is difficult to reliably predict what a user will do next after viewing a web page.
Lazy-loading of images is the opposite of pre-fetching content. Instead of loading images on the page, the browser defers loading of the resource until a later time. This technique is used to first populate only the content that appears “above the fold” – images that are outside of the viewport are not loaded immediately. If a user navigates away from a page prior to the image loading, this technique can also serve to reduce overall server load. Predicting what is in the viewport, though, is not always successful, resulting in images that should load not being visible.
Low Performance, High Reliability
Optimizations in this category provide lower performance benefits but are generally considered safe. Reasons for lower performance benefits are that the optimizations only apply to a small subset of users or to a small number of resources. Domain sharding was popular back in the day when browsers only opened 2 connections per domain; as browsers have evolved, they now support more than 6 connections per domain. For those still forced to use IE 6, domain sharding is still an excellent solution to performance issues. As with other techniques, HTTP/2 will eliminate the effectiveness of domain sharding.
Minification complements compression, making text files even smaller by eliminating whitespace and comments, as these aren’t needed by the browser to render a page. Minification can provide small improvements for already-compressed content and greater compression for users who can’t receive compressed content. With a number of tools out there to minify content, the time requirements are relatively small, but the reliability is high. Implementing minification makes sense if you have already optimized everything else on your site.
Figure 2: After minification
Revisiting the earlier concatenation example, we can see the benefits that minification can provide. While concatenation made the performance worse, minifying the content reduces page load time by 12% from the origin and 16.8% from the concatenated version, as the 600-Kb JS file is reduced to 428 Kb.
High Performance, High Reliability
This last category is where you want to focus your attention, as you get the most bang for your buck – the greatest performance gains and the lowest chance of breaking your application. Caching and compression are two very easy ways to improve performance and can generally be configured easily at the server or application delivery controller (ADC). Caching can also be extended a step further by using a content delivery network (CDN) to cache and serve content for geographically-distributed locations.
Many FEO techniques focus on improving the performance of first-time visitors, as the hope is that on a repeat visit, content will come from cache. The challenge is that not all caches are created equal.
Image optimization provides tremendous gains to web sites but requires a large time commitment to ensure that it is done correctly. Image optimization includes both lossless and lossy optimizations, such as converting an image from one format to another, stripping metadata, and reducing image quality. When done incorrectly, these can be highly unreliable and result in a broken site. Taking the time to optimize images correctly is required to get the high performance impact and high reliability for your application.
Where to start?
Optimizing applications seems a lot like a science experiment. You form a hypothesis that doing x will improve performance. You make the changes, realize something went wrong, investigate what went wrong, fix the problem, and test again. You may or may not get the performance improvements you were hoping for, and if you had estimated the project would take a couple of days and ended up taking weeks, was it worth it? If you are short on time, focus on the high-performance and high-reliability items with short times to implement, like caching and compression. Or stay tuned for the next blog post on how Instart Logic took a different approach to optimize application delivery by harnessing the power of machine learning.