OpenPOWER Summit North America 2019

In Spring of 2019, Virginia Tech hosted an OpenPOWER Hackathon with the aim of showing users how the PowerAI ecosystem can accelerate AI/DL frameworks with little to no modifications.  Many of the students that signed up for the Hackathon were new to Linux and HPC.  As a limited duration Hackathon with a goal of using the PowerAI toolset, it was important to get new users into the Power system quickly.  Here we will discuss the Open OnDemand platform that allows for creating custom GUIs and portals of entry into HPC systems.  We will show how this system allowed users to focus more on the Hackathon topic and give summaries of the winning projects and acceleration results achieved. 

UNICAMP (State University of Campinas) is an important OpenPower academic member providing community services such as Minicloud (http://openpower.ic.unicamp.br/minicloud/) that is one of the most widely used "free" developer clouds in the OpenPower ecosystem, providing virtual images hosting to dozens of Open-Source projects, academic students and faculty, as well as local IBM partners working on porting their solutions to POWER. This session will bring an update about our OpenPOWER Lab@Unicamp activities, including some statistics of usage such as geography, distro flavor, technology themes, etc. This session will also update the community on the current work to use the minicloud as a CI validation for multiple projects and will collect feedback and input on what other projects might be good targets to bring into minicloud, including a possible partnership with other foundation members.

Mellanox continues leadership as the best interconnect worldwide for both HPC and AI.  With the latest intelligent interconnect devices, such as Mellanox Quantum HDR 200Gb/s and ConnectX-6 are enabling the next generation of in-network co-processing.  With a more effective mapping of communication between devices in the system dramatically increases system performance.

Mellanox solutions are being used today in 59% of the HPC systems listed on the June 2019 TOP500 list, including the top three supercomputers in the world, and four of the top five in the list.  InfiniBand accelerates the fastest HPC and AI supercomputer in the world – the Oak Ridge National Laboratory ‘Summit’ system, delivering 148.6 Petaflops, and the second world’s fastest “Sierra” at Lawrence Livermore National Laboratory, demonstrating the robust ecosystem supported by the OpenPOWER Foundation and its membership.

Oregon State University’s Center for Genome Research and Biocomputing (CGRB) works closely with its users to determine the best computing platform for computational research. In the past several years we have seen many differences associated with architecture and GPU interactions. As we continued our testing we found tremendous benefits from using the OpenPOWER platform with Power9 processors and NVIDIA GPU. In an effort to truly determine the best platform for different types of GPU processing we created a set of test.  We will be showing all the current data comparing some of the most bleeding-edge platforms and a demo of the machines racing each other. Our goal is to provide users with the ability to bring online the best platform for the work they are planning with GPU technology. 

Fast Radio Bursts (FRBs) are momentary pulses of intense radio emissions which blast for a few milliseconds. The origin and source of FRBs is not yet identified, which make it a very exciting area of study. Due to their short-lived nature, real-time detection and analysis of FRBs become extremely difficult. One of the major and compute-intensive steps in this is radio frequency interference (RFI) calculation.

Our project under the OpenPOWER academia aims to optimize the performance of the RFI mitigation computations using OpenMP and CUDA on high-performance CPUs and GPUs. The performance and accuracy of the pipeline is currently being tested on OpenPOWER system with one Nvidia V100 and Power 9 CPU, the results and improvements in performances will be shared. Feather testing and optimization on multi GPU environments should increase the performance significantly.

The Oregon State University Open Source Lab (OSUOSL) works closely with Free and Open Source Software (FOSS) projects to provide them with a stable environment to improve their OpenPOWER porting needs. Since 2011, we’ve deployed and managed an OpenStack cluster powered by POWER9 and POWER8) used by over 100 projects. In this talk we will describe the technical challenges we have faced from setting up the initial cluster, to managing new users’ needs, tooling involved for managing the system, upgrading OpenStack and its components and other related topics. In addition, we’ll discuss a continuous integration platform we deployed which includes access to Nvidia GPU hardware along with a few other new services that are upcoming.

There are 200 billion stars in the Milky Way. The European Space Agency's Gaia project has mapped about 1% of them. The stars now need to be classified in much the same way as we classify animals and plants. Atos working with the University of Geneva has investigated the use of machine learning to help with this classification problem. Jez's talk describes the Gaia project before presenting the approach taken to construct and optimise a neural network capable of classifying the different star types. The techniques used are applicable to many different supervised machine learning problems. The talk will conclude with the initial results and a look towards future work.

This talk will provide an overview of the existing power measurement, power management, and performance-boosting capabilities of the POWER architecture.  Various levels of the hardware/firmware and software stack will be described.  Then, it will focus on a new data collection method enabled by OpenBMC that provides a rich data set including performance metrics, temperatures, and power consumption.  This new technology has been deployed at scale at ORNL with excellent results.  Additionally, a newly emerging power/performance-optimizing application, called GEOPM, will be described.