automatically exploiting cross-invocation parallelism using runtime ...

3.5 Related Work3.5.1 Cross-invocation ParallelizationLoop fusion techniques [22, 76] aggregate small loop invocations into a large loop invocation,converting the problem of cross-invocation parallelization into the problem of crossiterationparallelization. The applicability of these techniques is limited to mainly affineloops due to their reliance upon static dependence analysis. Since DOMORE is a runtimetechnique, it is able to handle programs with input-dependent dynamic dependences.Tseng [72] partitions iterations within the same loop invocation so that cross-invocationdependences flow within the same working thread. Compared to DOMORE, this techniqueis much more conservative. DOMORE allows dependences to manifest between threadsand synchronizations are enforced only when real conflicts are detected at runtime.While manually parallelizing a sequential program, programmers can use annotationsprovided by BOP [19] or TCC [25] systems to specify the potential concurrent code regions.Those code regions will be speculatively executed in parallel at runtime. Bothtechniques can be applied to exploit cross-invocation parallelism. However, they requiremanual annotation or parallelization by programmers while DOMORE is a fully automaticparallelization technique.3.5.2 Synchronization OptimizationsOptimization techniques are proposed to improve the performance of parallel programswith excessive synchronizations (e.g, locks, flags and barriers).Fuzzy Barrier [24] specifies a synchronization range rather than a specific synchronizationpoint. Instead of waiting, threads can execute some instructions beyond the synchronizationpoint. Speculative Lock Elision [57] and speculative synchronizations [40] designhardware units to allow threads to speculatively execute across synchronizations. Grace [4]wraps code between fork and join points into transactions, removing barrier synchroniza-46