2014 Rice Oil & Gas HPC has ended
Back To Schedule
Thursday, March 6 • 4:30pm - 6:30pm
Poster: Performance Challenges for Emerging HPC Systems, Milind Chabbi, Rice University

Sign up or log in to save this to your schedule, view media, leave feedback and see who's attending!


Today’s supercomputers are complex and enormous in scale. As a result, harnessing their full power is exceedingly difficult. Application performance problems of interest on HPC systems include both node-level and system-wide performance issues. Node-level performance issues include utilization of memory hierarchies as well as instruction, vector, and/or thread-level parallelism, as well as power consumption. System-wide problems include load imbalance and serialization across nodes, as well as communication and I/O bottlenecks. Failure to avoid or tolerate these issues can lead to major system-wide performance bottlenecks at scale. Even worse, one application may experience performance problems that result from heavy resource consumption by other jobs. In the face of all of this complexity, tools are essential for identifying code regions that consume excessive resources (e.g., time or power) relative to the work they accomplish, quantifying their impact, and diagnosing the root causes of their inefficiency. Unique processors, memory hierarchies, accelerators, network topologies, and software stacks each require different tool support for measurement and analysis. Effective performance tools for today’s supercomputers require support ranging from hardware to application domain. To date, performance tools have focused on post-mortem analysis of application performance to pinpoint and resolve causes of performance losses. For exascale systems faced with scarce resources (especially power), efficient resource management will require programs, libraries, runtime systems, and the operating system to analyze their own performance on the fly and initiate changes, e.g., migration of work or frequency scaling, to reduce resource consumption or improve utilization. As a result, exascale systems will need new software support to analyze performance measurements on the fly and policies to determine how to react. Designing the necessary performance tools’ interfaces for measurement, analysis, and control, as well as the mechanisms to support them is a key ingredient for the success of exascale systems.


Thursday March 6, 2014 4:30pm - 6:30pm PST
BRC Exhibit Hall Rice University 6500 Main Street at University, Houston, TX 77030