SC17: What do we see coming from the Super Computer World?
January 17, 2018 | BY: Tammy Carter
There were numerous tutorials on OpenMP, MPI, and OpenACC, as well as associated tools. One of the advanced topics presented how to use tasking in OpenMP to optimize irregular data flows with runtime dependent execution, as well as overlapping I/O and computational tasks. MPI classes featured discussions on collectives and topologies for optimization. One of the common topics this year was the hybrid-programming model, using OpenMP and MPI in combination. The focus was on performance aspects such as data and thread locality on NUMA architectures, false sharing, cache utilization and vectorization.
Over on the network side, advanced features of Infiniband, Omnipath, High-Speed Ethernet (HSE) and RoCE were discussed, and the challenges of RDMA, performance, scalability, and portability. One of the new IB transports is the eXtended Reliable Connection (XRC), which connects between nodes instead of processes. Another hot topic is On Demand Paging (ODP), both implicit and explicit, which simplifies programming while optimizing physical memory optimization along with unlimited memory region sizes. Manipulation of data while being transferred in the switch network via MPI operations is being address by the Scalable Hierarch (SHARrP).
The increased emphasis on GPUs could be seen everywhere. NVIDIA was prominent on the floor, in the tutorial and lectures for both acceleration of applications and deep learning. For example, the top 15 HPC applications are now GPU accelerated. All of the major server vendors are offering Volta and Pascal GPUs.
A Bright Spot in Curtiss-Wright's OpenHPEC Tool Suite
Wednesday at SC17, I was privileged to give a talk at the Bright Computing Theater on how Curtiss-Wright is using the Bright Cluster Manager to reduce the cost and time to deployment of HPEC systems. The audience was excited, and in some cases surprised, to see how their world is extending into our embedded one.