NUMA-BTDM: A Thread Mapping Algorithm for Balanced Data Locality on NUMA Systems

Optimizing for Non-Uniform Memory Access (NUMA) systems could be considered inappropriate because hardware architecture aware optimizations are not portable. On the contrary, this paper supports the idea that developing NUMA aware optimizations improves performance and energy consumption on NUMA systems and that these optimizations may be considered portable when they are non static. This paper introduces NUMA Balanced Thread and Data Mapping (BTDM), an extension of PThreads4w API [1]. NUMA-BTDM employs balanced data locality concept, improving thread and data mapping for NUMA systems. The purpose is to combine task parallelism with balanced data locality in order to obtain both better performance and reduced energy consumption on NUMA systems at run-time. The implementation of NUMA-BTDM targets homogeneous architectures based on the energy model with constant energy consumption or on the energy model in which each core is powered from a separate source (architectures on which parallel execution may reduce energy consumption compared to serial execution).