Abstract

In this thesis we develop preload, a daemon that prefetches binaries and shared libraries from the hard disk to main memory on desktop computer systems, to achieve faster application start-up times. Preload is adaptive: it monitors applications that the user runs, and by analyzing this data, predicts what applications she might run in the near future, and fetches those binaries and their dependencies into memory. We build a Markov-based probabilistic model capturing the correlation between every two applications on the system. The model is then used to infer the probability that each application may be started in the near future. These probabilities are used to choose files to prefetch into the main memory. Special care is taken to not degrade system performance and only prefetch when enough resources are available. Preload is implemented as a user-space application running on Linux 2.6 systems.

Citations