Our work might be thought of somewhat like a deadline scheduler. Instead of deadlines, we focus on release-times--i.e. when the application would like to execute. In our model, the application partitions it's own work into time-critical and computational events. It uses a new system call to supply timing (release-times) directly to the kernel, and to cooperate with other time-sensitive tasks.
A central thesis of our work is that responsiveness (running when you want to run) and fairness (amount of time one gets) should be independent.
Bottom line: we can maintain 1ms range responsiveness even while the system is 100% loaded. Our work is targeted to adaptive applications, i.e. those that can adapt quality by skipping less important work, so as to maintain timing for the more important stuff.
In the paper, we give examples with video, and also modified the X11 server.