Optimization techniques for adaptability in MPI application

Resumo

The first version of MPI (Message Passing Interface) was released in 1994. At that time, scientific applications for HPC (High Performance Computing) were characterized by a static execution environment. These applications usually had regular computation and communication patterns, operated on dense data structures accessed with good data locality, and ran on homogeneous computing platforms. For these reasons, MPI has become the de facto standard for developing scientific parallel applications for HPC during the last decades. In recent years scientific applications have evolved in order to cope with several challenges posed by different fields of engineering, economics and medicine among others. These challenges include large amounts of data stored in irregular and sparse data structures with poor data locality to be processed in parallel (big data), algorithms with irregular computation and communication patterns, and heterogeneous computing platforms (grid, cloud and heterogeneous cluster). On the other hand, over the last years MPI has introduced relevant improvements and new features in order to meet the requirements of dynamic execution environments. Some of them include asynchronous non-blocking communications, collective I/O routines and the dynamic process management interface introduced in MPI 2.0. The dynamic process management interface allows the application to spawn new processes at runtime and enable communication with them. However, this feature has some technical limitations that make the implementation of malleable MPI applications still a challenge. This thesis proposes FLEX-MPI, a runtime system that extends the functionalities of the MPI standard library and features optimization techniques for adaptability of MPI applications to dynamic execution environments. These techniques can significantly improve the performance and scalability of scientific applications and the overall efficiency of the HPC system on which they run. Specifically, FLEX-MPI focuses on dynamic load balancing and performance-aware malleability for parallel applications. The main goal of the design and implementation of the adaptability techniques is to efficiently execute MPI applications on a wide range of HPC platforms ranging from small to large-scale systems. Dynamic load balancing allows FLEX-MPI to adapt the workload assignments at runtime to the performance of the computing elements that execute the parallel application. On the other hand, performance-aware malleability leverages the dynamic process management interface of MPI to change the number of processes of the application at runtime. This feature allows to improve the performance of applications that exhibit irregular computation patterns and execute in computing systems with dynamic availability of resources. One of the main features of these techniques is that they do not require user intervention nor prior knowledge of the underlying hardware. We have validated and evaluated the performance of the adaptability techniques with three parallel MPI benchmarks and different execution environments with homogeneous and heterogeneous cluster configurations. The results show that FLEXMPI significantly improves the performance of applications when running with the support of dynamic load balancing and malleability, along with a substantial enhancement of their scalability and an improvement of the overall system efficiency.