This talk will describe different usage models for MCDRAM in Intel Xeon Phi processors (Knights Landing) as well as OpenMP affinity control for Intel Xeon Scalable processors (Skylake). The STREAM benchmark and the Parallel Research Kernels will be used to show how performance (especially memory bandwidth) depends on the different settings. The results demonstrate the trade-offs made between performance and programmability resulting from how high-bandwidth memory is exposed to applications. The talk will also cover the importance of handling affinity properly when using threads in a NUMA environment.

About the Speaker

Jeff Hammond is a System Architect at Intel, where his focus is hardware-software co-design in the context of exascale numerical simulations. He contributes to the development of open standards for parallel programming, including MPI, OpenMP, OpenSHMEM, and ISO C++. Previously, Jeff was a computational scientist at Argonne (ALCF), where he did things related to MPI. His academic background is in computational chemistry. See https://github.com/jeffhammond/ for details.