Evaluating techniques for parallelization tuning in MPI, OmpSs and MPI/OmpSs

Parallel programming is used to partition a computational problem among multiple processing units and to define how they interact (communicate and synchronize) in order to guarantee the correct result. The performance that is achieved when executing the parallel program on a parallel architecture is usually far from the optimal: computation unbalance and excessive interaction among processing units often cause lost cycles, reducing the efficiency of parallel computation. In this thesis we propose techniques oriented to better exploit parallelism in parallel applications, with emphasis in techniques that increase asynchronism. Theoretically, this type of parallelization tuning promises multiple benefits. First, it should mitigate communication and synchronization delays, thus increasing the overall performance. Furthermore, parallelization tuning should expose additional parallelism and therefore increase the scalability of execution. Finally, increased asynchronism would provide higher tolerance to slower networks and external noise. In the first part of this thesis, we study the potential for tuning MPI parallelism. More specifically, we explore automatic techniques to overlap communication and computation. We propose a speculative messaging technique that increases the overlap and requires no changes of the original MPI application. Our technique automatically identifies the application’s MPI activity and reinterprets that activity using optimally placed non-blocking MPI requests. We demonstrate that this overlapping technique increases the asynchronism of MPI messages, maximizing the overlap, and consequently leading to execution speedup and higher tolerance to bandwidth reduction. However, in the case of realistic scientific workloads, we show that the overlapping potential is significantly limited by the pattern by which each MPI process locally operates on MPI messages. In the second part of this thesis, we study the potential for tuning hybrid MPI/OmpSs parallelism. We try to gain a better understanding of the parallelism of hybrid MPI/OmpSs applications in order to evaluate how these applications would execute on future machines and to predict the execution bottlenecks that are likely to emerge. We explore how MPI/OmpSs applications could scale on the parallel machine with hundreds of cores per node. Furthermore, we investigate how this high parallelism within each node would reflect on the network constraints. We especially focus on identifying critical code sections in MPI/OmpSs. We devised a technique that quickly evaluates, for a given MPI/OmpSs application and the selected target machine, which code section should be optimized in order to gain the highest performance benefits. Also, this thesis studies techniques to quickly explore the potential OmpSs parallelism inherent in applications. We provide mechanisms to easily evaluate potential parallelism of any task decomposition. Furthermore, we describe an iterative trialand-error approach to search for a task decomposition that will expose sufficient parallelism for a given target machine. Finally, we explore potential of automating the iterative approach by capturing the programmers’ experience into an expert system that can autonomously lead the search process. Also, throughout the work on this thesis, we designed development tools that can be useful to other researchers in the field. The most advanced of these tools is Tareador – a tool to help porting MPI applications to MPI/OmpSs programming model. Tareador provides a simple interface to propose some decomposition of a code into OmpSs tasks. Tareador dynamically calculates data dependencies among the annotated tasks, and automatically estimates the potential OmpSs parallelization. Furthermore, Tareador gives additional hints on how to complete the process of porting the application to OmpSs. Tareador already proved itself useful, by being included in the academic classes on parallel programming at UPC.