Restoration of Legacy Parallelism: Transforming Pthreads into Farm and Pipeline Patterns

Parallel patterns are a high-level programming paradigm that enables non-experts in parallelism to develop structured parallel programs that are maintainable, adaptive, and portable whilst achieving good performance on a variety of parallel systems. However, there still exists a large base of legacy-parallel code developed using ad-hoc methods and incorporating low-level parallel/concurrency libraries such as pthreads without any parallel patterns in the fundamental design. This code would benefit from being restructured and rewritten into pattern-based code. However, the process of rewriting the code is laborious and error-prone, due to typical concurrency and pthreading code being closely intertwined throughout the business logic of the program. In this paper, we present a new software restoration methodology, to transform legacy-parallel programs implemented using pthreads into structured farm and pipeline patterned equivalents. We demonstrate our restoration technique on a number of benchmarks, allowing the introduction of patterned farm and pipeline parallelism in the resulting code; we record improvements in cyclomatic complexity and speedups on a number of representative benchmarks.

TAPP: DNN Training for Task Allocation through Pipeline Parallelism Based on Distributed Deep Reinforcement Learning

The rapid development of artificial intelligence technology has made deep neural networks (DNNs) widely used in various fields. DNNs have been continuously growing in order to improve the accuracy and quality of the models. Moreover, traditional data/model parallelism is hard to expand due to communication bottlenecks and hardware efficiency issues. However, pipeline parallelism trains multiple batches, reducing training overheads, so that it can achieve better acceleration effect. Considering the complexity of solving the pipeline parallel task allocation problem in heterogeneous computing resources, in this paper, a task allocation in pipeline parallelism (TAPP) based on deep reinforcement learning, is proposed. In TAPP, the predictive network is trained by a policy gradient until it obtains the optimal pipeline parallel task allocation scheme and speeds up the model training. Experimental results show that, on average, the single-step training time of TAPP is decreased by 1.37 times and the proportion of communication time is reduced by 48.92%, compared with the data parallelism, bulk synchronous parallel (BSP).