Maurice Herlihy

(Brown University)

"Is Transactional Memory Scalable?"

(Vortrag im Rahmen der "Distinguished Lecture Series Spring 2008" des Max Planck Instituts für Software-Systeme)

While transactional memory promises to ease the task of programming emerging multicore architectures, questions remain concerning how well it scales to long transactions and many cores. In this talk, we identify identify two substantial limitations in the way current proposals handle synchronization and recovery. Synchronization is typically based on read/write conflicts: two transactions conflict if they access the same object (or location) and one access is a write. Recovery is (with some exceptions) typically all-or-nothing: a transaction either commits, and installs its changes, or aborts, and discards its changes. We argue that read-write synchronization and all-or-nothing recovery are not well-suited to environments with long-lived transactions, substantial contention, or both.
We describe ongoing research on how Transactional Memory can be extended to alleviate these obstacles to scalability. We describe how to exploit semantic knowledge to enhance concurrency, and how a checkpoint/continuation style of programming can support fine-grained recovery, and a novel application of Bloom filters to detect and avoid deadlocks.
Joint work with Eric Koskinen.

Zeit: Mittwoch, 18. Juni 2008, 16:00 Uhr
Ort: Saarbrücken, Gebäude MPI-SWS, Raum 019
Hinweis: Der Vortrag wird live an die TU Kaiserslautern Gebäude 49 Raum 204-206 übertragen.