Abstract

Transactional memory is an idea from back in 1995 that has recently become popular because of the move towards multicore architectures. It promises to make fine grained parallelism available at no extra programming cost to single threaded serial algorithms. However some problems with TM prevent its easy adoption. For one, privatization is a significant overhead in some circumstances. It also does not co-exist with legacy lock based access. Finally it does not allow irrevocable actions such as IO within a transaction.

In this talk I’ll cover some very preliminary ideas towards solving some of these problems. I will present a scalable lock design that can be acquired both in transactional and non-transactional mode and give examples that should serve to highlight how this can make TM more applicable to real world problems. Finally I will show an interesting application of scalable locks to build data structures that are perfectly concurrent and provide fair access semantics to threads (each thread will ultimately get to complete its access within a bounded number of critical section completions by other threads).