Presentation

DescriptionPower and energy are critical Exascale constraints. Achieving system efficiency is a challenge under such constraints due to dynamic application phase behavior, increasing variation in processor power efficiency resulting from manufacturing, and due to complexities arising from upcoming heterogeneous architectures. In order to address some of these challenges, Intel introduced GEOPM, an open-source, portable, hierarchical job-level runtime system to optimize for time-to-solution by leveraging techniques from learning and control systems. Other processor architectures like IBM POWER and ARM have recognized the value of GEOPM, and are currently working to add their specific processor support. In this hands-on, half-day tutorial, we will discuss the state-of-the-art power management techniques used by Intel, IBM, and ARM, and discuss GEOPM features and usability. A high-level overview of the GEOPM architecture, a walkthrough of the GEOPM plugin infrastructure, and use cases will be discussed. Third-party plugins developed as part of Exascale Computing Project’s GEOPM plugins will also be demoed. Algorithms used to speed up application critical path (through adaptive configuration selection) and mitigate process variation (through intelligent power management) will be discussed. State-of-the-art and best practices for leveraging power management techniques in modern processors will also be covered.

Content Level 85% of the tutorial content will be targeting typical beginner-to-intermediate HPC system users. (50%Beginners 35%Intermediate, 15%Advanced)

Target AudienceThree broad categories of audience: (a) HPC end-users who wish to improve the energy-efficiency of their applications while using power-constrained systems (b) System programmers who wish to contribute design different power management algorithms (c) System administrators and facilities managers that need to understand and optimize the data center for power-efficiency

PrerequisitesFor the hands-on portion of the tutorial will require the attendees to use their laptops and will rely on a strong internet connection that would enable them to connect (using ssh) to a remote test-bed cluster over a VPN. Instructions regarding connecting to a remotely hosted VPN server will be distributed to the attendees during the tutorial. The attendees are also encouraged to bring their own HPC benchmarks/codes/application if they wish to evaluate the extent of efficiency gains achievable while using GEOPM.

Authors