Concurrency Analysis for Sequential Consistency in Titanium

As the limits of uniprocessor machines are being approached, application writers and system vendors alike have been turning to multiprocessor machines for performance. The major CPU manufacturers all have recently or will shortly introduce chips with multiple cores. Such systems, along with traditional multiprocessor machines, allow all processors to simultaneously access shared memory. This requires a memory consistency model to be defined, which determines in what order the memory updates on one processor appear to the other processors. The memory consistency model can be specified at the level of a programming language, which is crucial for languages that can be used on a wide variety of hardware. Language designers have traditionally been very reluctant to use the simplest model, sequential consistency, in which memory operations appear to occur in the order specified in the original program. This reluctance is due to a perception that such a model incurs prohibitive performance penalties, since it prevents reordering of operations and requires memory fences to be inserted in order to force the underlying hardware to respect ordering. Language designers instead have used complicated models [15, 19] that aren’t well-understood by programmers, or worse, ill-defined [8]. This is very problematic for programmers, since many common techniques such as spin-locks and presence bits depend on the details of the memory consistency model in order to function correctly. Various techniques have been proposed in order to decrease the cost of sequential consistency. In this paper, we present an interprocedural concurrency analysis for the Titanium programming language that can increase the precision of one such technique, cycle detection. We present both a basic algorithm and a modified one that only considers program execution paths that can occur in practice and prove that both algorithms are correct. We then apply these algorithms to a set of benchmarks, showing that they are effective in reducing the number of fences required to enforce sequential consistency in most of the benchmarks.