Resumen de Atomic dataflow model
Vladimir Gajinov
With the recent switch in the design of general purpose processors from frequency scaling of a single processor core towards increasing the number of processor cores, parallel programming became important not only for scientific programming but also for general purpose programming. This also stressed the importance of programmability of existing parallel programming models which were primarily designed for performance. It was soon recognized that new programming models are needed that will make parallel programming possible not only to experts, but to a general programming community. Transactional Memory (TM) is an example which follows this premise. It improves dramatically over any previous synchronization mechanism in terms of programmability and composability, at the price of possibly reduced performance. The main source of performance degradation in Transactional Memory is the overhead of transactional execution. Our work on parallelizing Quake game engine is a clear example of this problem. We show that Software Transactional Memory is superior in terms of programmability compared to lock based programming, but that performance is hindered due to extreme amount of overhead introduced by transactional execution. In the meantime, a significant research effort has been invested in overcoming this problem. Our approach is aimed towards improving the performance of transactional code by reducing transactional data conflicts. The idea is based on the organization of the code in which highly conflicting data is promoted to dataflow tokens that coordinate the execution of transactions. The main contribution of this thesis is Atomic Dataflow model (ADF), a new task-based parallel programming model for C/C++ that integrates dataflow abstractions into the shared memory programming model. The ADF model provides language constructs that allow a programmer to delineate a program into a set of tasks and to explicitly define data dependencies for each task. The task dependency information is conveyed to the ADF runtime system that constructs a dataflow task graph that governs the execution of a program. Additionally, the ADF model allows tasks to share data. The key idea is that computation is triggered by dataflow between tasks but that, within a task, execution occurs by making atomic updates to common mutable state. To that end, the ADF model employs transactional memory, which guarantees atomicity of shared memory updates. The second contribution of this thesis is DaSH - the first comprehensive benchmark suite for hybrid dataflow and shared memory programming models. DaSH features 11 benchmarks, each representing one of the Berkeley dwarfs that capture patterns of communication and computation common to a wide range of emerging applications. DaSH includes sequential and shared-memory implementations based on OpenMP and TBB to facilitate easy comparison between hybrid dataflow implementations and traditional shared memory implementations. We use DaSH not only to evaluate the ADF model, but to also compare it with other two hybrid dataflow models in order to identify the advantages and shortcomings of such models, and motivate further research on their characteristics. Finally, we study applicability of hybrid dataflow models for parallelization of the game engine. We show that hybrid dataflow models decrease the complexity of the parallel game engine implementation by eliminating or restructuring the explicit synchronization that is necessary in shared memory implementations. The corresponding implementations also exhibit good scalability and better speedup than the shared memory parallel implementations, especially in the case of a highly congested game world that contains a large number of game objects. Ultimately, on an eight core machine we were able to obtain 4.72x speedup compared to the sequential baseline, and to improve 49% over the lock-based parallel implementation based on work-sharing.
