| This thesis addresses two research problems. The first problem is how to run demanding applications on mobile computers. Mobile computers, such as smart phones and personal digital assistants, have recently become more prevalent. However, the size of these mobile computers constrain the processing power, battery capacity, and memory they can carry. Therefore, it is difficult to run resource-intensive applications on these computers.;This thesis explores utilizing publicly-available compute servers, which we call surrogates, to augment mobile computers to remotely execute resource-intensive applications. I built a software system named Slingshot to demonstrate this concept. Slingshot replicates application state on surrogates and the user's home machine by encapsulating the application state in a virtual machine. The replica on the surrogate provides good response time as it is co-located with the mobile computer, and the replica on the home machine can provide data safety if a surrogate fails.;The second problem this thesis addresses is how to make configuration management easier for users. Software applications provide many configuration options that allow users to customizing them. However, software applications depend on shared libraries and configuration data and interact with other software applications through various communication channels. Such complex software dependencies make it difficult to configure software applications correctly.;I built a software system named AutoBash to automate many configuration management tasks for users, such as troubleshooting misconfigurations and running regression tests. AutoBash automatically finds a solution that transforms the system into a healthy state by using a set of predicates that test the system state to verify each solution. AutoBash leverages OS-level speculative execution to try many solutions and causal information tracking to reduce the time to run regression tests.;To reduce the user-effort required to write predicates in AutoBash, this thesis explores automatically generating predicates by observing the actions of ordinary users fixing configuration problems. The main results were that: (1) my method can infer predicates for all configuration bugs studied with very few false positives, (2) the majority is usually right, so using multiple traces can improve results. |
