Atomicity provides strong guarantees against errors caused by unanticipated thread interactions, but is difficult for programmers to implement with low-level concurrency primitives. With the introduction of multicore processors, the problems are compounded. Atomic sections are a high level language feature that programmers can use to designate the blocks of code that need to be free from unanticipated thread interactions, letting the language implementation handle the low-level details such as deadlock. From a language designer's point of view, the challenge is to implement atomic sections without compromising performance.
We propose an implementation of atomic sections that transparently inserts locking into object-oriented programs. The main advantages of our approach are: (1) We infer path expressions (that at run-time resolve to actual objects) for many more accesses in the atomic section than previous work could infer. (2) We use multi-granularity locking for guarding iterative traversals. (3) We ensure freedom from deadlock by rolling back the lock acquisition phase. (4) We release locks as early as possible. In summary, our approach uses a finer-grained locking discipline than previous lock inference techniques.
pubs.doc.ic.ac.uk: built & maintained by Ashok Argent-Katwala.