Static Lock Allocation

The allocation of lock objects to critical sections in concurrent programs affects both performance and correctness. Traditionally, this allocation is done manually by the programmer. Recent work explores automatic lock allocation, aiming primarily to minimize conflicts and maximize parallelism by allocating locks to individual critical sections. We investigate several modes of lock allocation, using connected components (groups) of interfering critical sections on a critical section interference graph as the basis for allocation decisions. Our allocator uses thread-based side effect analysis which is built from several pluggable component analyses. It benefits from precise points-to and may happen in parallel information. Thread-local object information provides a small improvement over points-to analysis alone. Our framework minimizes the restrictions on input programs, dealing gracefully with nesting and deadlock, and requiring only simple annotations identifying critical sections. Legacy programs using synchronized regions can be processed without alteration. We find that dynamic locks do not broadly improve upon identical allocations of static locks, but allocating several dynamic locks in place of a single static lock can significantly increase parallelism in certain situations. We experiment with a range of small and large Java benchmarks on 1 to 8 processors, and find that a singleton allocation is sufficient for five of our benchmarks, and that a static allocation with Spark points-to analysis is sufficient for another two. Of the other five benchmarks, two require the use of all phases of our analysis, one depends on using the lockset allocation, and two benchmarks proved too complex to be automatically transformed to satisfactory performance.