Lucas Leandro Nesi

Abstract

HPC platforms are vastly heterogeneous because of intra-node resources like accelerators and inter-node heterogeneity when there are different machines. The applications that use these resources are already very complex, with many distinct operations and phases, and developers must consider all sets of diverse computational resources. The task-based programming paradigm is a modern alternative to increase the computational efficiency of intra-node heterogeneous resources while maintaining relative development simplicity. The application defines a Direct Acyclic Graph of tasks and a dynamic runtime asynchronously schedules them to the resources respecting task dependencies. However, handling different types of nodes requires new specific strategies to distribute an application in this asynchronous and heterogeneous environment. This thesis studies the problem of distributing this type of complex task-based applications over those diverse system-level resources, proposing strategies to divide their load correctly, considering computational heterogeneity, multiple-phase asynchronism, and adaptability. This work uses real applications to validate its results with experiments conducted in large testbeds and a supercomputer. The thesis' main contributions are the following. (i) Strategies for distributing a single application operation considering the trade-off of communication, critical path, and heterogeneous load balancing. (ii) A set of optimizations for improving asynchronous phase overlap in applications. (iii) A methodology for computing the relative power of each phase on each heterogeneous group of nodes considering the phase overlap. (iv) A distribution strategy for an antecedent phase reducing communication redistribution. (v) A strategy for the application dynamically adapts during execution to decide the best subset of nodes for each phase. (vi) An extended comprehensive analysis of the experiments that include a methodology to analyze the application progress per node resilient to heterogeneity and that can cluster nodes with similar behavior. Ultimately, this thesis is a step toward efficiently exploiting and combining any of these diverse resources, using them to handle applications' distinct necessities better, and improving their overall performance.