Generational GC always better?

I have heard a lot of stories about different ways of implementing GCs in the past few years. Academic papers that are usually the base for such opinions are highly outdated and a real comparison of different GC implementations for all the different architectures that Firefox is running on has never been published. (Please please tell me if I am wrong and there is something that compares GC approaches that work all the way between high end desktop machines and low end smart phones!)

Here at Mozilla we have a cool animation that uses HTML 5 and WebGL features.
The Flight of the Navigator definitely shows every flaw of the GC because long GC events result in annoying pauses during the animation.
So lets compare the generational-moving approach that Google Chrome uses to our old-school "mark and sweep" approach.



The first graph shows the GC pause times for Chrome in msec. The second graph shows the same benchmark with a nightly build of Firefox.
Chrome has many GC events that are close to 4 msec but there are also the outliers that cause GC pauses of up to 900ms. We have constant pauses of around 100ms.
Our pause time is not perfect but comparing the video experience between the 2 browsers I definitely favor many barely noticeable 100ms pauses instead of some 900ms pauses where I start worrying if my browser has crashed.

We invested a lot of time decreasing the GC pause time. Our compartment approach assures that our GC has a minimal workload and we moved most of the sweeping cost off the critical path to a background thread.

We also go towards a generational and incremental GC model but we have to make sure that we watch our outliers closely.