CSR:Small:System Support for Petabyte Memories

University Of Wisconsin-Madison, Madison WI

Investigators

Abstract

Computers have long separated working memory from permanent storage. Memory has been managed via fixed-sized pages (like a page in a book) while storage gets managed with large, variable-sized files (like a whole book). Especially for computers that provide online services, memory is growing to mammoth size. Equally important, new technologies will allow memory to be permanent like storage. This project embraces these changes and will develop techniques to manage memory more like files to speed access to huge data sets in a mammoth amount of memory. The new techniques will be driven with new tools, data, and analysis, all made publicly available. This project has five thrusts. First, the project will build and distribute new tools for emulating computers with mammoth memories via virtualization and application-specific compression. Second, the project will develop "file-only memory" to support efficient management of the arbitrarily large regions that compose a mammoth memory. Third, the project will analyze and mitigate memory fragmentation that plagues old and new solutions to memory management. Fourth, the project will development a full software-hardware stack for address translation that scales efficiently to mammoth memory sizes. Fifth, the project will engage undergraduates in both teaching of computer science generally, and in the proposed research as assistants to graduate student researchers. This project promises more efficient computers with mammoth memories, leading to higher performance and fewer wasted resources. Through the development and release of emulation tools, this work will allow rapid efficient studies of mammoth memories not currently feasible. The study of fragmentation will spur further research on new memory management techniques to specifically address fragmentation of mammoth memories. The proposed system enhancements will enable mammoth-memory systems at reduced cost and increased efficiency and simplify management of large data sets. Finally, the investigators will release all workloads, tools, and emulation platforms as open source. The project repository will be available from the research group home page (http://research.cs.wisc.edu/multifacet/) as well as the investigators' home web pages. Emulation tools will be published to source code repositories allowing collaborative development. Data sets will be published in full on a web site hosted at the University of Wisconsin at Madison or be provided via a query interface to large data sets. Papers and articles will also be available on the research group web page. This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.