Last week we broke the story of RoboHornet, a new Web browser performance benchmark. Today, we chronicle the fallout of what is perhaps the worst-received benchmark launch in history, plus add the results of Microsoft's counter-move: RoboHornet Pro.
Holy train wreck, Batman! The end of September was one weird, wild week for Web browser news.
Despite RoboHornet being an independent GitHub project on paper, Microsoft and others aren't buying it. In possibly the fastest turnaround in development history, Microsoft has taken RoboHornet from first alpha release to "Pro" in less than 24 hours. In another surprising turn, Mozilla formally concurs with Redmond that RoboHornet's technical merits are questionable. Perhaps arch-rivals Microsoft and Mozilla finally realize that Google is their real enemy. After all, while the browser war veterans were busy battling each other, Chrome just waltzed through the front gates and usurped the throne.
Let's go through the wacky events of last week as they occurred:
Tuesday: Microsoft says "Thanks, but no thanks" to RoboHornet, dismissing the new test as a meaningless micro-benchmark that does not reflect real-world performance. Below is an excerpt from the IE Blog, and from the language used in the opening sentence, you can tell that Microsoft isn't buying the "independent benchmark" part:
Yesterday Google released its latest micro-benchmark, RoboHornet, in which Internet Explorer 10 scores rather well. While we appreciate the gesture, members of our engineering team took a look at the benchmark and found that RoboHornet isn’t all that representative of the performance users might encounter on real-world sites. Like all micro-benchmarks, RoboHornet is a lab test that only focuses on specific aspects of browser performance.
We decided to take the RoboHornet micro-benchmark and run it in the context of a real-world scenario. Using modern browser capabilities like CSS3 Animations, CSS3 Transforms, CSS3 Text Shadows, custom WOFF fonts, Unicode, Touch, and more, we created a site that looks a little bit like the familiar Matrix. We then ran the RoboHornet micro-benchmark in the context of this real website. While running both the Matrix and RoboHornet micro-benchmark at the same time, Chrome slows to a crawl and stops animating the screen, because it wasn’t designed to handle a benchmark load in the context of a real-world scenario. Meanwhile Internet Explorer 10 remains responsive, continues animating the screen, and finishes the test in less than half the time that Chrome does… We have made RoboHornet Pro available on IE Test Drive, so you can check it out for yourself.
Later that day Mozilla's Justin Lebar opened a bug on RoboHornet's GitHub page entitled Eliminate and outlaw micro-benchmarks where he states:
If you guys want us (in my case, Mozilla) to take robohornet seriously, I strongly recommend you write some macrobenchmarks and eliminate the microbenchmarks from your test suite.
Wednesday: Microsoft-affiliated steward, John David Dalton, drops all mention of Microsoft from his RoboHornet stewardship. Meanwhile, the Mozilla steward, Daniel Buchner, leaves the committee entirely.
Today: We have Mozilla's official response to the RoboHornet debacle:
There are a lot of benchmarks out there and different benchmarks for the same particular task can behave very differently. Many benchmarks are self-serving, in that the creators will typically pick out a set of programs that they think are worthwhile to get faster on and then only after turning these workloads share them with the wider community. However, what developers and browser implementers really need to have here is good benchmarks that allow us to better see holistic performance. We'd like to see more benchmarks created that focus on the entire consumer experience--for example, benchmarks that focus on interaction with the browser, popular web applications or sites, and common tasks like panning and zooming on mobile.
Micro-benchmarks, like RoboHornet, do not accurately reflect the user experience on the Web. RoboHornet aims to measure real performance, but it falls short. RoboHornet is currently 17 micro-benchmarks, each of which measures one thing that a website can do. But real websites do hundreds and thousands of things, so almost all of them are entirely unaccounted for by RoboHornet. RoboHornet lists some things that are currently slow in browsers and points them out for attention from vendors. That's a useful service, but it's not the same as accurately representing real performance.
It is still unclear what exactly Mozilla thinks about Microsoft's RoboHornet Pro, but now that results are published, we'll ask for a statement and update the story accordingly.
Of the remaining top five browser vendors, Opera refuses to comment, and Apple cannot be reached.
RoboHornet Pro is already off to a better start than RoboHornet because it works with so many more browsers than the original test. We are now able to include results from the Windows 8 Metro versions of Chrome and IE10, as well as Maxthon and Sleipnir on both Windows and OS X. iOS browsers can now run the test as well. We've included Yahoo! Axis, Google Chrome, Dolphin, Mozilla Firefox, Maxthon, Apple Safari, and Sleipnir. While Android loses its stock browser, it retains Chrome and gains Opera Mobile.
Let's quickly recap the test setup, and then see how all these browsers fare on RoboHornet Pro.