Memory Footprint of Locality Information on Many-Core Platforms

Exploiting the power of HPC platforms requires knowledge of their increasingly complex hardware topologies. Multiple components of the software stack, for instance MPI implementations or OpenMP runtimes, now perform their own topology discovery to find out the available cores and memory, and to better place tasks based on their affinities. We study in this article the impact of this topology discovery in terms of memory footprint. Storing locality information wastes an amount of physical memory that is becoming an issue on many-core platforms on the road to exascale. We demonstrate that this information may be factorized between processes by using a shared-memory region. Our analysis of the physical and virtual memories in supercomputing architectures shows that this shared region can be mapped at the same virtual address in all processes, hence dramatically simplifying the software implementation. Our implementation in hwloc and Open MPI shows a memory footprint that does not increase with the number of MPI ranks per node anymore. Moreover the job launch time is decreased by more than a factor of 2 on an Intel Knights Landing Xeon Phi and on a 96-core NUMA platform.