Achieving transparency mapping parallel applications
Edgar A. León, Matthieu Hautreux · Proceedings of the International Symposium on Memory Systems · 2018
Computer systems are becoming increasingly complex: they provide the expected compute capability at the cost of deeper memory hierarchies (high-bandwidth and high-capacity memories), heterogeneous compute elements (latency-optimized and throughput-optimized cores), and heterogeneous memories (volatile and non-volatile). To run a parallel application, users need to determine the mapping of MPI tasks and OpenMP/POSIX threads to the hardware resources. Not only this can be challenging but when executing the same application on a different system, the mapping will likely change to attain reasonable performance.