This project focuses on an new approach to multiprocessor synchronization called transactional memory. This approach replaces conventional locking techniques with light-weight transactions: sequences of instructions that ap-pear to execute atomically. Transactional memory relies on a combination of hardware and software support. The project explores the interface between hardware and software support for transactions. Hardware is well-suited for short transactions that access a small memory region, but is less effective for longer transactions, or transactions that access more memory. Transactions that cannot be executed directly in hardware must be executed partially or en-tirely in software, and the project is investigating a range of possible trade-offs. The project also explores the implications of the transactional memory model for the design of highly-concurrent data structures. While transactional mem-ory provides better support for fine-grained concurrency than conventional ock-ing techniques, it also provides opportunities for new kinds of data structure designs and algorithms.
This project addresses a fundamental problem in modern computing. Every year, processors get faster and cheaper. Recently, however, fundamental limita-tions such as heat dissipation have made advances in clock speed increasingly difficult. An alternative approach, called multiprocessing, harnesses multiple processors to work together. Unfortunately, current techniques for multiproc-essor synchronization are difficult to use, and scale poorly to large-scale sys-tems. This alternative synchronization model investigated by this project prom-ises to one substantial barrier to more effective multiprocessor computing. The project's results will be disseminated through scholarly publication and open-source software.
Dr. Brett D. Fleisch Program Director, CISE/CNS May 26, 2004. .
