Introducing Support for Brotli Compression
This post was written by our software developer intern Denys Tsomenko, who worked on a Brotli compression library during his internship.
Currently, ASP.NET developers have two compression methods available to use in their web applications: Deflate and gzip. But there is a trade off between compression time and size reduction. Different algorithms can perform quite differently. In 2015, two engineers at Google designed a new compression algorithm called Brotli that can have a better compression without spending more time. Brotli is already supported by the most browsers such as Google Chrome, Mozilla Firefox, Opera, and Microsoft Edge.
In this post, we’ll take a look at how Brotli performs and showcase our early alpha preview for it.
The quality of every compression algorithm depends on three main factors:
- Compression Ratio is the main value for these algorithms, because we generally want to reduce size as much as possible.
- Compression Time is important because it describes how much the compressor has to spend. It’s especially important for web servers where CPU time tends to be precious.
- Decompression Time is important because it controls how much time the consumer has to spend for reading compressed data. Again, this can also impact server CPU load as servers are frequently clients in micro-service architectures.
Compression ratio indicates how much a compression algorithm can reduce the size and is computed as follows:
Compression Ratio := Uncompressed Size / Compressed Size
So the higher the number, the better the algorithm is at compressing data. To judge how well Brotli does here, we’ll look at various different data sets.
Compression algorithms usually allow for tuning how much time the algorithm will spend to compress the data. In .NET, we expose this via the CompressionLevel enumeration. To make numbers comparable, we use the extremes, which are
Fastest, which means we let the algorithm compress as fast as possible, with
Optimal, which means we let the algorithm spend as much time as it wants.
Let’s start by looking at a typical file format that occurs in web development: CSS. For our particular sample, we can see that Brotli compression ratio is better on
Optimal and about the same on
The first three columns show size reduction with
Fastest, the next three with
Optimal and the last one with a middle quality level for Brotli. As you can see in the graph above, even at middle quality level, Brotli compression ratio is higher than the optimal quality level of both gzip and Deflate.
To assess whether Brotli can also provide savings outside of the realm of web files, let’s take a look at the Canterbury Corpus, which is popular set of files for testing compression algorithms:
The files were chosen because their results on existing compression algorithms are “typical”, and so it is hoped this will also be true for new methods. — Canterbury Corpus
Specifically, we’ll look at these files:
|ptt5||CCITT test set||513,216|
Looking at the data, it’s clear that Brotli is particular great for larger files, but even for smaller files Brotli is either on par or slightly better than Deflate and gzip:
As explained in the introduction, compression is particular interesting for web scenarios to reduce load times. However, if the compression algorithm takes too long, than those savings are immediately lost again. Thus, it’s also important to compare how fast an algorithm can compress.
The graph below highlights the difference in compression times between Brotli, Deflate, and gzip. We used a larger file (around 4 MB) to measure the compression time. Since we measure time, lower is better.
Fastest, Brotli is faster than both Deflate and gzip:
However, when set to
Optimal, Brotli takes a lot more time:
So this tells us that the time Brotli spends can significantly differ depending on the quality level. To understand where the sweet spot it, let’s just focus on Brotli and how the time changes based on the compression levels that Brotli supports, which is a range from 1 to 11. Using a level above 7 means you’ll have to accept orders of magnitude differences in compression time, so for cases where time is important, you should probably use lower values. However, as we’ve seen earlier, even at level 5, Brotli compresses quite well.
For static files, you might not care about compression time as you only have to do it once. However, in virtually all scenarios the client will have to decompress on every request, which makes decompression speed quite important.
Choosing the same 4 MB file for testing we get the following results:
As you can see, Brotli is again much faster than gzip and quite close to the performance of Deflate (it’s worth pointing out that for smaller files, decompression speed is comparable across all algorithms).
When the Brotli compression algorithm was released in 2015, the TV show Silicon Valley was popular, and characters used the Weissman score to measure the performance of their compression algorithm. While the Weissman score was developed for fictional use on that show, it’s quite useful because it provides a handy efficiency metric.
So let’s calculate it for Brotli: we’ll use gzip as the baseline and Brotli as the comparison. The algorithm considers both, compression ratio and time spent. A score of less than 1 means that gzip is better while a score greater than 1 indicates Brotli fares better. As you can see, for every file we tested Brotli is better than gzip based on this metric:
Today, .NET support for Brotli compression is available as an alpha-quality preview. You can find the source code in our CoreFxLab repository.
If you want to give it a try, follow these steps:
- Create a new project. Brotli compression is available for .NET Standard 1.4, so either a .NET Framework or a .NET Core application will work.
- Register our MyGet feed with the preview builds. Please note that you shouldn’t register this feed globally as it might result in bringing in preview packages into your production projects. Instead, register this feed only for the projects you want to use preview bits in. You can can do that by creating a file called
nuget.configand putting it next to your solution file:
- Install the System.IO.Compression.Brotli package. Make sure you check Include prerelease in the NuGet package manager UI.
With this setup, let’s take a look at the source you’d write when compressing and decompressing with Brotli. Similar to the existing compression, we offer a
BrotliStream that allows you to compress and decompress data via streams.
For compression, this would look as follows:
Decompression works analogously except that you pass in
CompressionMode.Decompress to the constructor:
Now let’s see how we can start using Brotli for response compression in an ASP.NET Core application. We’ll do that by extending the Response Compression middle-ware.
Start by creating a web application. I’ve used ASP.NET Core 2.0 with the MVC template.
Create a new a file called
Open the file
Startup.cs and modify
ConfigureServices to add the Brotli compression provider:
To enable automatic response compression, you also need to modify
Configure by adding this line:
To see it in action, launch the app in your favorite browser (I’m using Google Chrome here). To make things more interesting, I’m simulating a typical 3G connection via the F12 developer tools:
Results with Brotli:
Results without Brotli:
As you can see, we reduced the data from 219 KB to 180KB (~18%) and the page load time from 1.33s to 1.12s (~16%).
Brotli is a relatively new compression algorithm. It’s quite beneficial for web clients, with decompression performance comparable to gzip while significantly improving the compression ratio. These are powerful properties for serving static content such as fonts and html pages. Thus, if you currently use gzip or Deflate as compression for your web site (or have never used compression) you should give Brotli a try.
An early alpha preview version of the Brotli compression library is available for you today. You can use the preview in both .NET Framework and .NET Core applications. In typical sites we’ve seen faster load times between 14% to 30%, but your mileage will vary.
Please give us feedback to make Brotli more usable and efficient! Let us know what you think by leaving a comment here or by filing an issue in the CoreFxLab repository.