The new Traverse supercomputer, which is composed of 46 IBM POWER9 nodes with 4 NVIDIA V100 GPUs per node, has an impressive peak performance of over 1.4 PFLOPS. However, to take full advantage of this computational power, one must have specialized knowledge of both the hardware and software. This workshop will show participants how to leverage the advanced capabilities of Traverse including GPU Direct, NVLink, CUDA Multi-Process Service (MPS), CUDA-Aware MPI, NVIDIA V100 GPU Tensor Cores, Scalable Hierarchical Aggregation Protocol (SHArP), IBM POWER9 architecture, VSX vectorization, Simultaneous Multithreading (SMT), GPU-enabled IBM Engineering and Scientific Subroutine Library (ESSL), IBM XL and PGI compilers, and NVMe storage.
IMPORTANT: “All visitors (including Gabe and yourself) will need to be known in advance in order to come on site at PPPL. No big deal for US citizens. Foreign nationals will need to fill out some extra paperwork though.”
Learning objectives: Attendees will learn how to use the advanced features of Traverse
Knowledge prerequisites: Background in Linux clusters and/or supercomputers
Hardware/software prerequisites: None
Workshop format: Lecture and demonstration
Please register online at princeton.edu/training