Structuring the execution of OpenMP applications for multicore architectures

Abstract

International audienceThe now commonplace multi-core chips have introduced, by design, a deep hierarchy of memory and cache banks within parallel computers as a tradeoff between the user friendliness of shared memory on the one side, and memory access scalability and efficiency on the other side. However, to get high performance out of such machines requires a dynamic mapping of application tasks and data onto the underlying architecture. Moreover, depending on the application behavior, this mapping should favor cache affinity, memory bandwidth, computation synchrony, or a combination of these. The great challenge is then to perform this hardware-dependent mapping in a portable, abstract way. To meet this need, we propose a new, hierarchical approach to the execution of OpenMP threads onto multicore machines. Our ForestGOMP runtime system dynamically generates structured trees out of OpenMP programs. It collects relationship information about threads and data as well. This information is used together with scheduling hints and hardware counter feedback by the scheduler to select the most appropriate threads and data distribution. ForestGOMP features a high-level platform for developing and tuning portable threads schedulers. We present several applications for which we developed specific scheduling policies that achieve excellent speedups on 16-core machines

A Sharing-Aware Memory Management Unit for Online Mapping in Multi-core Architectures

“…These mechanisms generate mapping information based on a very small number of samples compared to SAMMU, as all memory accesses are handled by the MMU. Some techniques such as Forest-GOMP [4] require annotations in the source code and depend on specific parallelization libraries. Similarly, Ogasawara [20] proposes a data mapping method that is limited to object oriented languages.…”

Scalable Task Parallelism for NUMA

“…ForestGOMP [5,6] is an OpenMP run-time with a resourceaware scheduler and a NUMA-aware allocator. It introduces three concepts: grouping of OpenMP threads into bubbles, scheduling of threads and bubbles using a hierarchy of runqueues, and migrating data dynamically upon load balancing.…”

“…On the operating system side, optimizations are compelled to place tasks and data conservatively [13,24], unless provided with detailed affinity information by the application [5,6], high-level libraries [26] or domain specific languages [20]. Nevertheless, as task-parallel run-times operate in user-space, a separate kernel component would add additional complexity to the solution; this advocates for a user-space approach.…”

Hardware-Assisted Thread and Data Mapping in Hierarchical Multicore Architectures

“…A library called ForestGOMP is introduced in [Broquedis et al 2010a]. This library integrates into the OpenMP runtime environment and gathers information about the different parallel sections of the applications.…”