Performance Analysis Tools Working Session

What are our application performance goals for exascale? What questions can performance analysis tools realistically answer (i.e., what roles will performance analysis play)?
What context/semantics are needed to improve our ability to attribute behavior to root causes?
What data are we most interested in, what interfaces (HW or SW) are needed to gather that data, and at what levels?
How do we best incorporate our analysis tools into applications to gather this data?
Given our analysis goals, what are the best methods to address scalability, how might hierarchical analysis play a role, and what assistance is needed from other HW/SW layers?
Beyond the usual software engineering concerns, what components are common across multiple performance analysis tools, or shared with other system infrastructure?
Is there a role for machine learning in postmortem tools and for the analysis of data collected by an application when it runs on “performance collection mode"?

Is auto-tuning a tool or as a component of a compiler/runtime system? Is it part of a performance tool? Conversely do tools have a place inside applications?
Do we need management tools (e.g., power management, fault management, etc.)? Or are the management functions part of a software stack component (e.g. runtime system)? Same question regarding optimization: does it belong in tools, or does it belong in compilers, runtime systems, etc?
At Exascale with continually evolving HW (due to fault tolerance), does auto-tuning ever stop or is it always going on alongside the application?
What has and has not worked in design, deployment, and training for previous tools to promote their widespread adoption, use, and impact?
What elements of popular/enduring tools can be extended to or reproduced at Exascale? What lessons are there from tools that survived the tera-to-petascale transition?
How to measure and reward developing “industrial strength” tools? What should be the role of open source in this process?

Highlights