Personal tools
You are here: Home Workshops Summer 2008 Workshop on Automatic Tuning for Petascale Systems
Document Actions

Workshop on Automatic Tuning for Petascale Systems

by admin last modified 2008-07-30 04:14 — expired

July 8-11, 2008, Snowbird Ski & Summer Resort, Snowbird, Utah, USA


  • Kathy Yelick (University of California at Berkeley), "yelick"  AT ""

  • Jack Dongarra (University of Tennessee), “dongarra” AT “”

  • Rich Vuduc (Georgia Institute of Technology), “richie” AT “”

  • Keith Cooper (Rice University), "keith" AT ""


Over the last decade, microprocessor features such as deep pipelines, multiple cores, and complex memory hierarchies have made it increasingly difficult to achieve good performance in scientific applications and libraries. This fact has given rise to systems that automate the tuning process, using large amounts of computation to configure the application or library for good performance on the target architecture.

This workshop, the second in the series, will bring together researchers and practitioners in automatic tuning, in library design and construction, and in compiler-based code generation to identify and discuss opportunities and challenges in the use of automatic tuning for future petascale systems. As with last year, we are actively soliciting members of the compiler and library autotuning communities. In addition, we are interested in attracting attendees whose focus is on runtime optimization of scientific programs.

The workshop will consist of a series of talks and discussion sessions. We anticipate that most workshop attendees will present a talk on their research, their insights, and their experience. The discussion sessions will focus on opportunities to build shared infrastructure, along with issues raised during the workshops.


The emphasis of this year's workshop is on the boundaries and interactions between specialized, library-based autotuning systems and current and future compilation environments. In particular, we ask the speakers to present research talks that take a "position" on one or more of the following questions (alternatively, feel free to make up your own). It would be helpful if your first slide explicitly stated the questions/positions you are taking. We will use the open discussion periods and working dinners to tackle these questions as well.

Positions and Questions

  • Proposition: Today's autotuning work does/does not address the challenges of petascale. (What challenges need to be addressed?)
  • Question: How do we measure success for tuning? How do we measure and weigh performance? Productivity?
  • Question: What architectures/platforms should we target?
  • Proposition/Question: “Parameter tuning” is the wrong focus for our area, and will lead only to incremental improvements; what problems should we look at instead?
  • Proposition: Self-tuned libraries will always outperform compiler-generated code.
  • Question: What improvements should we expect from autotuning, at both the compiler level and the library level? Will we allow future compilers to change data structures and algorithms?
  • Proposition: Simple performance models (e.g., a cache-oblivious model, or assuming systems with simple in-order cores) will be the right models in the future and will obviate the need for empirical search.
  • Proposition: The traditional boundaries between applications, libraries, compilers, and operating systems is too rigid and needs to be changed. Example: What in our app-specific schedulers could be moved into the OS, if anything?
  • Question: What issues are we as a community ignoring? What other technologies should we investigate to find application-specific, platform-specific improvement?
  • Question: What can we do to build common tool bases for compiler-based autotuning and for construction of self-tuning or autotuning libraries?
  • Proposition: The focus on specialized tuning systems is too narrow, and so only compilers, which apply most broadly, are the most sensible investment.
  • Proposition: Runtime optimization will catch opportunities for improvement that neither a compiler nor a neither an autotuned library can.
  • Question: Suppose all layers of the software stack (e.g., OS, middleware, MPI, libraries, apps) are “autotuned.” Will we need to integrate these multiple layers, and if so, how?

Day 1 - Tuesday July 8

  • 6:00-8:00pm: Working buffet dinner.
    Discussion topics: autotuning opportunities, successes to date, remaining challenges, and promising approaches.

Day 2 - Wednesday July 9

Afternoon Evening
  • 6:00-8:00pm: Working buffet dinner.
    Discussion topics: deriving application parameters, application characterization with performance counters, using regression models to help select optimizations, and plans for moving forward with autotuning.

Day 3 - Thursday, July 10

  • Lunch / social activity


This workshop was sponsored by the Center for Scalable Application Development Software, with funding from the DOE Office of Science through the Scientific Discovery through Advanced Computing (SciDAC) program.

« September 2017 »
Su Mo Tu We Th Fr Sa

Powered by Plone

CScADS Collaborators include:

Rice University ANL UCB UTK WISC