Friday, February 21, 2014

SIAM 2014 Minisymposium: Visualization of Performance Data on Large Scale Systems and Applications

Minisymposium 74

4:50 PM - 6:30 PM
Room: Salon A

The hardware complexity of HPC systems has increased in parallel with the complexity of modern HPC applications, which has made writing efficient software difficult. Understanding the interactions between hardware and software and their impacts at large scale is essential for optimizing HPC systems, but results from classic performance tools are often too low-level and difficult to comprehend. To overcome this challenge and to gain real insight into an application’s performance, we need both novel techniques in performance analysis as well as a close collaboration between the fields of performance analysis and data analytics/visualization, which this minisymposium aims at fostering.


Organizers:

Martin Schulz
Lawrence Livermore National Laboratory, USA

Joshua Levine
Clemson University, USA

Peer-Timo Bremer
Lawrence Livermore National Laboratory and University of Utah, USA

Paul Rosen
University of Utah, USA


Speakers:

	Allen Malony, University of Oregon, USA (PRESENTATION) Title: Visions of Tau Dancing in Your Head - Ruminations on Performance Visualization Abstract: Modern performance tools can capture arbitrarily complex information about the execution of parallel applications. Making "sense" of the performance data is as much a process of analysis, to extract high value content based on knowledge of what is "meaningful" about performance, as it is of presentation, to convey the relevant features of the results in forms that enable "understanding" and "insight" for users. The talk explores ideas for performance visualization in the TAU Performance System.
	Kwan-Liu Ma and Chris Muelder, University of California, Davis, USA (PRESENTATION) Title: Visual Characterization of High-End Computing  Abstract: Modern supercomputers are complex, hierarchical systems consisting of huge numbers of cores, systems for disk storage, and nodes for I/O forwarding. These numbers continue to grow and the need for tools to understand the behavior of both the applications and system software becomes paramount. This talk presents visual analytics techniques that address the challenge of understanding the behavior of complex software on very large-scale compute platforms, like the current petascale computers, and shows the resulting tools can potentially tune that software to attain the highest possible degrees of efficiency.
	Peer-Timo Bremer, Lawrence Livermore National Laboratory and University of Utah, USA (PRESENTATION) Title: Visual Performance Analysis for the Exascale Era Abstract:  Performance analysis of parallel scientific codes is becoming increasingly difficult due to the rapidly growing complexity of applications and existing tools fall short in providing intuitive views to reveal the root causes of performance problems. We have developed a new paradigm of projecting and visualizing performance data obtained from one domain onto other domains using projections and implemented in our BoxFish tool set
	Nathan Tallent, Pacific Northwest National Laboratory, USA (PRESENTATION|SLIDES) Title: Effectively Presenting Application Performance Using Simple Techniques Abstract:  Call path profiling and tracing have been shown to provide insight into the performance characteristics of complex parallel programs. However, poor presentation of performance data obscures insight. To enable rapid analysis of an execution's performance bottlenecks, we describe a small set of simple presentation techniques. Using these simple techniques in concert has proven to be very effective. These techniques form the basis of the HPCToolkit performance tools.