Cloud computing enables flexible, dynamic outsourcing while improving cost efficiencies. These operational and economic benefits, however, are today not available to safety-critical and mission-critical real-time applications due to the lack of timeliness supports by current cloud infrastructures. Providing real-time guarantees on cloud platforms faces several novel challenges because cloud platforms are not designed for response time guarantees but to achieve elasticity and the illusion of abundant resources available on demand.
This research extends current cloud infrastructures with resource management techniques to enable real-time guarantees on the cloud. The proposed work focuses on the scheduling of data-intensive real-time applications onto cloud resources to meet their timing requirements. The project (i) proposes a formal framework for modeling and performance evaluation of real-time applications in cloud environments, with a focus on data-parallel middleware; (ii) develops algorithms for scheduling continuous streams of real-time cloud applications in an online setting; (iii) designs techniques to address issues introduced by virtualization and machine failures based on probabilistic models, hierarchical scheduling techniques, multi-mode techniques, and feedback control techniques; and (v) evaluates these techniques in the context of practical real-time applications.
The research will result in the development of a real-time variant of at least one data-parallel middleware infrastructure by enhancing existing open-source platforms. Research prototypes will be released as open source to spur the research and development of cloud computing targeting real-time applications and systems. Research concepts and tools developed in this project will be incorporated into relevant courses at Penn.