Human-Centered Scheduling of Interactive and Multimedia Applications on a Loaded Desktop

Desktop operating systems such as Windows or Linux base scheduling decisions on the CPU usage of competing processes. Processes that exhibit low CPU usage are given priority, under the assumption that they are interactive and spend most of their time waiting for user input. However, this doesn’t work for modern multimedia applications, which require significant CPU resources, so playing a DVD movie while running a heavy background job will usually result in poor graphical quality. We therefore suggest a new way to identify interactive processes, by explicitly measuring direct interactions with the user. For example, for each process we measure the rate at which it draws pixels on the screen. Given this data, CPU resources can be allocated preferentially to processes that perform user I/O. A scheduler following these principles has been implemented in a Linux environment, and measurements indicate that it is very effective in distinguishing between competing interactive and non-interactive processes, and allocating CPU accordingly. It also handles competition between multiple interactive jobs by allocations that lead to equal I/O rates. This is completely automatic and transparent to the different applications, but requires some system modifications.