Automatic task and data mapping in shared memory architectures

Abstract

Reducing the cost of memory accesses, both in terms of performance and energy consumption, is a major challenge in shared-memory architectures. Modern systems have deep and complex memory hierarchies with multiple cache levels and memory controllers, leading to a Non-Uniform Memory Access (NUMA) behavior. In such systems, there are two ways to improve the memory affinity: First, by mapping tasks that share data (communicate) to cores with a shared cache, cache usage and communication performance are improved. Second, by mapping memory pages to memory controllers that perform the most accesses to them and are not overloaded, the average cost of accesses is reduced. We call these two techniques task mapping and data mapping, respectively. For optimal results, task and data mapping need to be performed in an integrated way. Previous work in this area performs the mapping only separately, which limits the gains that can be achieved. Furthermore, most previous mechanisms require expensive operations, such as communication or memory access traces, to perform the mapping, require changes to the hardware or to the parallel application, or use a simple static mapping. These mechanisms can not be considered generic solutions for the mapping problem. In this thesis, we make two contributions to the mapping problem. First, we introduce a set of metrics and a methodology to analyze parallel applications in order to determine their suitability for an improved mapping and to evaluate the possible gains that can be achieved using an optimized mapping. Second, we propose two automatic mechanisms that perform task mapping and combined task/data mapping, respectively, during the execution of a parallel application. These mechanisms work on the operating system level and require no changes to the hardware, the applications themselves or their runtime libraries. An extensive evaluation with parallel applications from multiple benchmark suites as well as real scientific applications shows substantial performance and energy efficiency improvements that are significantly higher than simple mechanisms and previous work, while maintaining a low overhead.