Workload Clustering for Increasing Energy Savings on Embedded MPSOCS

Özcan Öztürk, Mahmut Kandemir, Sri Hari Krishna NarayananOpenalex Unknown 2012 · DOI · OpenAlex

Abstract

Voltage/frequency scaling and processor low-power modes (i.e., processor shut-down) are two important mechanisms used for reducing energy consumption in embedded MPSoCs. While a unified scheme that combines these two mechanisms can achieve significant savings in some cases, such an approach is limited by the code parallelization strategy employed. In this paper, we propose a novel, integer linear programming (ILP) based workload clustering strategy across parallel processors, oriented towards maximizing the number of idle processors without impacting original execution times. These idle processors can then be switched to a low power mode to maximize energy savings, whereas the remaining ones can make use of voltage/frequency scaling. Our simulation-based analysis reveals that the proposed ILP based approach is very effective in reducing the energy consumptions of the applications tested and generates much better energy savings than all the alternate schemes tested (including a unified scheme that combines voltage/frequency scaling and processor shutdown).

Introduction: The Energy-Performance Conflict

In modern embedded Multiprocessor Systems-on-a-Chip (MPSoCs), energy efficiency is as critical as performance. Traditionally, two main techniques are used to manage power: Voltage/Frequency Scaling (VFS), which slows down processors during light workloads, and Processor Shutdown, which turns off idle cores to eliminate leakage power.

However, standard compilers are optimized for performance, often spreading tasks across as many processors as possible to minimize execution time. This "wide and thin" distribution leaves many cores partially utilized—active enough to prevent shutdown, but idle enough to be inefficient. This paper addresses this gap by proposing a "workload clustering" strategy that consolidates tasks to maximize core idleness.

Methodology: Optimal Task Consolidation

The core of the proposed approach is an Integer Linear Programming (ILP) model that reassigns tasks to the minimum number of processors required to meet the original execution deadline ($T_{max}$). Unlike traditional parallelization, which seeks to balance load, workload clustering seeks to "pack" load.

Comparison of different power management strategies: (a) original workload movement, (b) voltage scaling, (c) shutdown, (d) simple unified, and (e) proposed clustering.

The ILP model ensures that every job runs exactly once, deadlines are met, and energy is minimized. Specifically, the objective function targets the sum of dynamic energy from active tasks and leakage energy from processors that remain powered on. The model also accounts for the energy overhead of potential cache misses caused by task migration, ensuring that the calculated savings are realistic.

Key Insights: The "Packing" Philosophy

The primary insight is that in sub-micron technologies (like the 70-nm process simulated), leakage energy is a dominant factor. Consolidating tasks onto fewer processors allows more cores to be completely shut down, which is far more effective than simply scaling the frequency of many active cores.

The final state of the proposed workload clustering, showing consolidated tasks and shut-down processors.

Once tasks are clustered, any remaining "slack" time on the active processors can still be exploited using VFS, providing a secondary layer of energy savings. This creates a "narrow and deep" workload profile that is inherently more energy-efficient than the traditional "wide and thin" profile.

Experimental Results

The researchers evaluated the clustering strategy using eight array-intensive embedded applications (including 3D modeling, digital filtering, and LU decomposition) on the SIMICS platform. The results were compared against three other common power management schemes:

  • Voltage Scaling (VS) only: ~6.55% average energy savings.
  • Shutdown (SD) only: ~7.36% average energy savings.
  • Simple Unified (VS + SD): ~13.52% average energy savings.
  • Proposed Clustering: ~22.65% average energy savings.

The clustering approach significantly outperformed the simple unified scheme, nearly doubling the savings in some cases. Furthermore, the benefits were found to scale positively with the number of available processors and voltage levels, making it a promising approach for future many-core architectures.

Conclusion

Workload clustering represents a paradigm shift in how we think about parallel task distribution for energy-constrained systems. By using ILP to optimally pack tasks, systems can significantly reduce leakage power by shutting down entire cores without sacrificing performance. While the computational complexity of solving ILP problems remains a challenge for real-time adjustments, the significant energy gains make it a compelling strategy for compiler-level optimizations in next-generation embedded MPSoCs.

Summary updated

Abstract

Voltage/frequency scaling andprocessor low-power modes (i.e., processor
\nshut-down) are two important mechanisms usedfor reducing energy
\nconsumption in embedded MPSoCs. While a unified scheme that combines
\nthese two mechanisms can achieve significant savings in some
\ncases, such an approach is limited by the code parallelization strategy
\nemployed. In this paper, we propose a novel, integer linear programming
\n(ILP) based workload clustering strategy across parallel processors,
\noriented towards maximizing the number of idle processors without
\nimpacting original execution times. These idle processors can then
\nbe switched to a low power mode to maximize energy savings, whereas
\nthe remaining ones can make use ofvoltage/frequency scaling. In order
\nto check whether this approach brings any energy benefits over the pure
\nvoltage scaling based, pure processor shut-down based, or a simple
\nunified scheme, we implemented four different approaches and tested
\nthem using a set of eight array/loop-intensive embedded applications.
\nOur simulation-based analysis reveals that the proposed ILP based approach
\n(1) is very effective in reducing the energy consumptions of the
\napplications tested and (2) generates much better energy savings than
\nall the alternate schemes tested (including a unified scheme that combines
\nvoltage/frequency scaling and processor shutdown).
BibTeX
@inproceedings{ozturk2012workload,
  author = {Özcan Öztürk and Mahmut Kandemir and Sri Hari Krishna Narayanan},
  title = {{Workload Clustering for Increasing Energy Savings on Embedded MPSOCS}},
  booktitle = {Openalex Unknown},
  year = {2012},
  doi = {10.1002/9781118342015.ch19},
}

Loading related content...

Figure