Discovery of Potential Parallelism in Sequential Programs

Abstract

In the era of multicore processors, the responsibility for performance gains has been shifted onto software developers. Once improvements of the sequential algorithm have been exhausted, software-managed parallelism is the only option left. However, writing parallel code is still difficult, especially when parallelizing sequential code written by someone else. A key task in this process is the identification of suitable parallelization targets in the source code. Parallelism discovery tools help developers to find such targets automatically. Unfortunately, tools that identify parallelism during compilation are usually conservative due to the lack of runtime information, and tools relying on runtime information primarily suffer from high overhead in terms of both time and memory. This dissertation presents a generic framework for parallelism discovery based on dynamic program analysis, supporting various types of parallelism while incurring practically affordable overhead. The framework contains two main components: an efficient data-dependence profiler and a set of parallelism discovery algorithms based on a language-independent concept called Computational Unit.The data-dependence profiler serves as the foundation of the parallelism discovery framework. Traditional dependence profiling approaches introduce a tremendous amount of time and memory overhead. To lower the overhead, current methods limit their scope to the subset of the dependence information needed for the analysis they have been created for, sacrificing generality and discouraging reuse. In contrast, the profiler shown in this thesis addresses the problem via signature-based memory management and a lock-free parallel design. It produces detailed dependences not only for sequential but also for multi-threaded code without causing prohibitive overhead, allowing it to serve as a generic base for various program analysis techniques.Computational Units (CUs) provide a language-independent foundation for parallelism discovery. CUs are computations that follow the read-compute-write pattern. Unlike other concepts, they are not restricted to predefined language constructs. A program is represented as a CU graph, in which vertexes are CUs and edges are data dependences. This allows parallelism to be detected that spreads across multiple language constructs, taking code refactoring into consideration. The parallelism discovery algorithms cover both loop and task parallelism.Results of our experiments show that 1) the efficient data-dependence profiler has a very competitive average slowdown of around 80× with accuracy higher than 99.6%; 2) the frame- I would also like to thank my master students and student assistants for their coding support.Wolfram Gottschlich implemented the original version of the signature described in this thesis.Tuan Dung Nguyen implemented the lock-free parallel version of the DiscoPoP profiler. MichaelBeaumont tested the memory skipping technique. Daniel Fried implemented the method of characterizing DOALL loops using machin...

Improving Performance of Transactional Applications through Adaptive Transactional Memory

“…The adaptive system relies on the prediction of the decision tree to select either HTM or STM. § Our evaluations using STAMP [2], NAS [4], and DiscoPoP [37] benchmark suites reveal that on average, the adaptive system is able to improve speed of transactional applications by 20.82%.…”

“…This benchmark suite was developed for a tool to automatically find potential parallelism in sequential programs [37]. This tool is called DiscoPoP which is able to find parallelism between code regions with subjective granularity.…”

“…The first set of experiments are based on evaluation of RTM and TinySTM on 12 benchmarks taken from Stamp [2], NAS [4], and DiscoPop [37] benchmark suites. This evaluation is primarily conducted to compare the performances of the two systems.…”

“…Our work is orthogonal and can be combined with this paper[3] to enhance performance further.Z. Li et al[37] presented a compiler-based tool, called DiscoPoP, to automatically identify regions of code that can be executed in parallel. It is designed to be able to find code regions with arbitrary granularity.…”

Predicting Parallelization of Sequential Programs Using Supervised Learning

“…We use the DiscoPoP (Discovery of Potential Parallelism) tool [10] to extract dynamic profiling and instruction dependency data from instrumented versions of sequential programs. First, sequential programs are compiled to the Low Level Virtual Machine (LLVM)'s intermediate representation (IR) [9].…”