Adapting Application Mapping to Systematic Within-Die Process Variations on Chip Multiprocessors

Yang Ding, Mahmut T. Kandemir, Mary Jane Irwin, Padma RaghavanHiPEAC 2009 · DOI

TL;DR

This paper tackles the challenge of performance and power variations across identical cores in modern chip multiprocessors (CMPs) caused by manufacturing process variations. They propose and evaluate methods to intelligently map applications to these cores, significantly improving energy-delay product (EDP) by up to 73.4% compared to naive approaches.

What Problem Does This Paper Address?

As semiconductor manufacturing processes shrink, it becomes increasingly difficult to ensure that all components on a chip are identical. This leads to "process variations," where even identically designed cores can have slightly different performance and power characteristics. These variations can be systematic (e.g., due to lens aberrations in lithography) or random. For chip multiprocessors (CMPs), which have multiple cores, these variations mean that some cores might be faster but consume more power, while others are slower but more power-efficient. Traditionally, designers have addressed this by assuming the worst-case scenario for all cores, leading to wasted resources and increased power consumption. This paper argues that software can and should be aware of these variations to optimize application execution. Specifically, they focus on "intra-die" variations (within a single chip) and how to adapt application mapping to these core-to-core differences to minimize the energy-delay product (EDP), a key metric balancing performance and energy efficiency.

Methods

The core of this work lies in developing intelligent thread-to-core mapping strategies. They propose two main approaches:

  1. Process Variation Aware Thread Mapping: This involves using profile data to understand how different applications perform and consume energy on cores with varying characteristics (bins). They then employ Integer Linear Programming (ILP) to find the optimal mapping of application threads to processor bins to minimize the EDP. This approach acknowledges that different applications might benefit from different types of cores (fast/leaky vs. slow/efficient).
  2. Changing Bin Frequencies: Building on the first approach, they explore dynamically adjusting the operating frequencies of groups of cores (bins). Instead of clocking all cores at the lowest common frequency, they allow certain bins to run at their higher native frequencies, while others might be "downgraded" to save power. This adds another dimension to the optimization problem, allowing for more fine-grained control over performance and energy. Again, ILP is used to find optimal configurations.

They also investigate both non-preemptive (applications use available cores without disturbing others) and preemptive (cores can be reallocated from existing applications for a new one) mapping strategies.

Datasets

The authors evaluate their methods using:

  • NAS Parallel Benchmark Suite 3.2: This includes serial (for single-threaded) and MPI/OpenMP implementations.
  • SPEC Benchmark Suites: Specifically, SPEC CPU 2006 for single-threaded applications and SPEC OMP 2001 for multi-threaded applications.

They use a simulator (Simics) to model a 16-core CMP architecture and employ power models (Wattch-like) and cache energy models (CACTI 5.0).

Key Insights and Ideas

  • Software Can Exploit Hardware Asymmetry: The paper strongly demonstrates that by making the software aware of process variations, significant performance and energy benefits can be achieved, contrary to the traditional hardware-centric "worst-case" approach.
  • Application Preferences Vary: Not all applications benefit from the fastest cores. Some are more sensitive to leakage power and perform better on slower, more energy-efficient cores, highlighting the need for tailored mapping.
  • Dynamic Frequency Scaling is Powerful: Allowing groups of cores (bins) to operate at different frequencies, rather than a uniform low frequency, offers substantial EDP savings.
  • ILP is a Scalable Optimization Tool: Integer Linear Programming proves to be an effective method for finding optimal thread-to-core mappings and frequency configurations, even with complex constraints.
  • Partial Profiling is Viable: The approach can still yield good results even with incomplete profile data, thanks to curve fitting techniques.

Key Results

  • Process Variation Aware Thread Mapping: Achieved an average EDP saving of 11.7% compared to a variation-agnostic scheme.
  • Changing Bin Frequencies: This approach, combined with variation-aware mapping, led to average EDP savings of 37.1%.
  • Preemptive Mapping: Provided an additional 37.1% extra EDP savings on average compared to non-preemptive mapping.
  • Maximum EDP Savings: In some cases, they achieved EDP savings as high as 73.4%.
  • ILP Solver Overhead: The overhead incurred by the ILP solver was found to be negligible.

Limitations and Future Work

  • Profile Data Dependency: While they show partial profiling works, the accuracy of the results still depends on the quality and quantity of profile data.
  • Preemptive Mapping Complexity: Implementing preemptive mapping in a real system involves complex thread migration and checkpointing mechanisms, which were not the focus of this study.
  • Systematic Variations Focus: The work primarily addresses systematic variations. The impact of random variations was not explicitly studied.
  • Future Work: The authors suggest extending the ILP formulation to handle more complex scenarios, such as dynamic thread migration and more sophisticated frequency/voltage scaling strategies. They also hint at the possibility of adapting their approach to other types of process variations.

Summary updated

BibTeX
@inproceedings{ding2009adapting,
  author = {Yang Ding and Mahmut T. Kandemir and Mary Jane Irwin and Padma Raghavan},
  title = {{Adapting Application Mapping to Systematic Within-Die Process Variations on Chip Multiprocessors}},
  booktitle = {HiPEAC},
  year = {2009},
  doi = {10.1007/978-3-540-92990-1\_18},
}
Figure