Adaptive Parallelism for Coupled, Multithreaded Message-Passing Programs

Hybrid parallel programming models that combine message passing (MP) and shared- memory multithreading (MT) are becoming more popular, especially with applications requiring higher degrees of parallelism and scalability. Consequently, coupled parallel programs, those built via the integration of independently developed and optimized software libraries linked into a single application, increasingly comprise message-passing libraries with differing preferred degrees of threading, resulting in thread-level heterogeneity. Retroactively matching threading levels between independently developed and maintained libraries is difficult, and the challenge is exacerbated because contemporary middleware services provide only static scheduling policies over entire program executions, necessitating suboptimal, over-subscribed or under-subscribed, configurations. In coupled applications, a poorly configured component can lead to overall poor application performance, suboptimal resource utilization, and increased time-to-solution. So it is critical that each library executes in a manner consistent with its design and tuning for a particular system architecture and workload. Therefore, there is a need for techniques that address dynamic, conflicting configurations in coupled multithreaded message-passing (MT-MP) programs. Our thesis is that we can achieve significant performance improvements over static under-subscribed approaches through reconfigurable execution environments that consider compute phase parallelization strategies along with both hardware and software characteristics. In this work, we present new ways to structure, execute, and analyze coupled MT- MP programs. Our study begins with an examination of contemporary approaches used to accommodate thread-level heterogeneity in coupled MT-MP programs. Here we identify potential inefficiencies in how these programs are structured and executed in the high-performance computing domain. We then present and evaluate a novel approach for accommodating thread-level heterogeneity. Our approach enables full utilization of all available compute resources throughout an application’s execution by providing programmable facilities with modest overheads to dynamically reconfigure runtime environments for compute phases with differing threading factors and affinities. Our performance results show that for a majority of the tested scientific workloads our approach and corresponding open-source reference implementation render speedups greater than 50 % over the static under-subscribed baseline. Motivated by our examination of reconfigurable execution environments and their memory overhead, we also study the memory attribution problem: the inability to predict or evaluate during runtime where the available memory is used across the software stack comprising the application, reusable software libraries, and supporting runtime infrastructure. Specifically, dynamic adaptation requires runtime intervention, which by its nature introduces additional runtime and memory overhead. To better understand the latter, we propose and evaluate a new way to quantify component-level memory usage from unmodified binaries dynamically linked to a message-passing communication library. Our experimental results show that our approach and corresponding implementation accurately measure memory resource usage as a function of time, scale, communication workload, and software or hardware system architecture, clearly distinguishing between application and communication library usage at a per-process level