PxP Runtime System image

Co-Managing Quality-Performance-Power Tradeoffs
Click here to access project overview slides.
Voltage Scaling [publication 1] image

Quality and Performance. Many large-scale scientific simulations require the numerical solution of models based on time-dependent nonlinear partial differential equations in two or three spatial dimensions. The nonlinear equations in these systems are often solved with implicit Newton-type methods or semi-implicit schemes, both of which leverage sparse or irregular computational kernels for basic operations, such as mesh management, differentiation, and sparse linear system solution. Each kernel has a multitude of implementations offering a wide range of tradeoffs in solution quality (e.g., accuracy, reliability, and scalability) and performance (e.g., execution time/rate and parallel efficiency/speedup). Thus, proper method selection to meet changing application quality-of-service requirements and changing technologies can potentially provide dramatic performance improvements. A principal challenge is to automate dynamic method selection; without automation, the problem of selecting and using the best solution method is impractical, and the application community cannot easily reap the benefits of research of the last several decades.

Power. Current high-end platforms are ensembles of multiple fast CPUs with deep memory hierarchies and high-speed interconnects. Geometric scaling of raw power (Moore's Law) arises from more and faster transistors on a chip. However, chips are approaching their packaging thermal limits, and the power-related costs for high-end systems, both electrical power consumed (in Megawatts) and machine room cooling loads (200 W/ft^2), continue to grow as a quadratic function of peak execution rates and clock frequencies. Although a faster scientific simulation, such as one obtained by exploiting quality-performance tradeoffs, is also often one that consumes less power by using fewer compute cycles, a major challenge is developing explicitly power-aware scientific computing tools. In the near term, such tools can deliver lower-power realizations without adversely impacting application performance, and in the longer term, provide insights that can be used by the designers of systems software and microprocessors to develop future systems.

Our project seeks to address these two challenges by developing adaptive software tools to co-manage quality-performance-power tradeoffs. Key aspects of our project (initiated in 9/04) include:

  • Combinatorial and statistical adaptive techniques to select methods dynamically, to deliver improved performance while producing a solution that meets application quality requirements.
  • Using an annotated model of computation and communication costs and sparse data access patterns, to develop techniques for power reduction without performance impairment (using techniques such as DVS).
  • Implementing our techniques by developing an adaptive component software system on high-end multiprocessors such as the Blue-Gene system.
  • Related Publications

  • 1. Reducing Power with Performance Constraints for Parallel Sparse Applications , G. Chen, K. Malkowski, M. Kandemir, and P. Raghavan, submitted to the International Parallel and Distributed Symposium, IPDPS05. Click here to access a copy of the paper (pdf).
    We reduce overall energy consumption by scaling the voltages/frequencies of those processors that are not in the critical path; i.e., our approach is oriented towards saving power without incurring performance penalties. Our experiments with matrices extracted from real applications indicate that our recursive strategies are very effective in saving power, and the savings achieved are close to the optimal values.
  • Participants

  • M. J. Irwin,   K. Malkowski   and P. Raghavan (Pennsylvania State University).
  • L. C. McInnes,   and B. Norris (Argonne National Laboratory).
  • Funding

  • National Science Foundation, "Adaptive Software for Extreme-Scale Scientific Computing: Co-Managing Quality-Performance-Power Tradeoffs", P. Raghavan, M. J. Irwin, L. C. McInnes, and B. Norris, 9/04--08/07.
  • Open Post-Doctoral Position

  • Note: We have an opening for a post-doctoral researcher with a Ph.D. in Computer Science and Engineering and a strong background in high-performance scientific computing. Interested candidates should contact P. Raghavan.

  • Return to Padma Raghavan's Webpage