A recent trend in the commercial, high-technology sector has been to make large-scale, proprietary computing and storage infrastructure available to the public. Companies such as Amazon, Google, and many others, enable access to these vast resource collections via the Internet, and provide a wide range of services for accessing them (tools, APIs, resource scheduling and accounting, etc.). Users are encouraged to think of the infrastructure as a an opaque computing ``cloud'' that delivers compute and storage capabilities in the same way that the network ``cloud'' delivers network connectivity: through a client-provider relationship governed by well-defined interfaces. Unfortunately, at present, cloud computing technologies provide support for only a limited set of applications from the web service domain. As a result, remotely managed web hosting is, by far, the most prevalent usage model for cloud systems.
With this research, the PIs propose to design, implement, and evaluate the extensions to the services offered by the two most prevalent cloud purveyors (Amazon.com and Google) to facilitate usage by a more broad developer base (e.g., scientists and students) who require support for general compute-intensive applications. The basis of this work will be two novel research projects that are able to implement commercial cloud computing functionality in academic settings: AppScale and Eucalyptus. The PIs will employ these infrastructures to investigate and contribute the three fundamental technologies that are missing from extant cloud technologies but that are necessary to enable support of high-performance, compute intensive applications within cloud-based clusters: efficient support of shared-memory inter-process communication, high-performance file system support, and system-wide, adaptive performance monitoring and analysis.
