Optimizing throughput/power trade-offs in hardware transactional memory using DVFS and intelligent scheduling

Clay Hughes, Tao Li2011 · DOI

TL;DR

This paper tackles the challenge of optimizing both performance and power consumption in modern multi-core processors that utilize Hardware Transactional Memory (HTM). The authors introduce novel techniques combining Dynamic Voltage and Frequency Scaling (DVFS) with an intelligent transaction scheduler to reduce energy consumption and improve throughput, showing significant gains over existing methods.

What Problem Does This Paper Address?

Modern processors are increasingly packed with more processing elements (PEs) to handle multithreaded workloads. However, efficiently exploiting these resources, especially with complex synchronization mechanisms like locks, is difficult for software developers. Hardware Transactional Memory (HTM) offers a promising abstraction to simplify parallel programming by managing synchronization automatically. Yet, HTM systems can suffer from performance bottlenecks due to transactional stalls and aborts, leading to increased power consumption and reduced throughput. This paper aims to bridge the gap between the growing need for energy efficiency and the potential of HTM by proposing and evaluating new power and performance optimization techniques specifically for HTM systems.

Methods

The paper proposes two main techniques to optimize HTM performance and power:

  1. Dynamic Voltage and Frequency Scaling (DVFS) during Transactional Stalls: This policy dynamically adjusts the voltage and frequency of individual PEs based on their transactional state. When a PE is stalled (e.g., waiting for a lock or due to a conflict), its frequency is scaled down to save power. Conversely, if a PE is actively trying to resolve a conflict or commit, its frequency might be scaled up to improve throughput. This aims to increase overall throughput and energy efficiency by better managing PE utilization.
  2. Intelligent Transaction Scheduling and Clock Gating: This approach uses a transaction's "conflict probability" (based on recent aborts and active transactions) to predict potential contention. If a transaction is deemed too likely to abort, it can be preemptively stalled. While stalled, the PE's clocks are gated (turned off) to drastically reduce dynamic power consumption. This reduces contention and the overhead associated with aborts.

These techniques are integrated into a cycle-accurate simulator that models a 4-core CMP system with HTM capabilities.

Datasets

The evaluation uses a combination of benchmark suites and synthetic workloads:

  • SPLASH-2 and STAMP benchmarks: These are standard suites used to evaluate parallel applications. The STAMP suite, in particular, is highlighted for its relevance to future transactional workloads.
  • Synthetic benchmarks: These were generated using the TransPlant tool to explore a wider range of transactional behaviors, especially focusing on varying transaction granularity. This allows for a more controlled analysis of how the proposed techniques perform under different conditions.

Key Insights and Ideas

  • Stalls and Aborts are Key Power/Performance Bottlenecks: The paper's analysis clearly shows that transactional stalls and aborts are significant contributors to both wasted power and reduced performance in HTM systems. Targeting these specific events is crucial for optimization.
  • DVFS Can Be Fine-Tuned for HTM: Applying DVFS not just based on overall system load but on the transactional state of individual PEs offers a more granular and effective way to manage power and boost performance.
  • Predictive Stalling Reduces Abort Overhead: By predicting potential conflicts and preemptively stalling transactions, the system can avoid costly aborts and the subsequent re-execution. This proactive approach is more efficient than reactive conflict resolution.
  • Combining DVFS and Scheduling Amplifies Benefits: The synergistic effect of DVFS (managing power during stalls) and intelligent scheduling (preventing stalls/aborts) leads to greater improvements than either technique alone.
  • Contention is the Differentiator: The effectiveness of the proposed techniques, especially the scheduling policy, is most pronounced in benchmarks with high contention, where aborts and stalls are frequent.

Key Results

  • Combined DVFS and Scheduling: The combined approach reduced the Energy-Delay Squared product (ED2P) by an average of 19% for eager-eager, 20% for eager-lazy, and 15% for lazy-lazy HTM configurations.
  • Synthetic Workloads: For synthetic workloads, the combined policies showed even greater improvements, reducing ED2P by an average of 29%.
  • Labyrinth Benchmark: This highly contentious benchmark saw significant improvements, with ED2P reductions of up to 76% for some configurations.
  • Comparison to Prior Work: The proposed methods showed improvements of 12-30% over previous techniques like clock gating and transactional serialization.

Limitations and Future Work

  • Overly Pessimistic Scheduling: For certain benchmarks (e.g., genome, raytrace), the conflict probability scheduler could be overly conservative, leading to increased execution time and thus higher ED2P, even if power was reduced. This suggests a need for more adaptive or per-transaction scheduling parameters.
  • Idealized Simulator: The simulator models some aspects (like clock gating wake-up) as instantaneous or very fast, which might not perfectly reflect real-world hardware latencies.
  • Future Work: The authors suggest exploring more sophisticated feedback mechanisms for the conflict prediction algorithm and investigating how these techniques can be applied to even larger many-core systems. They also note the potential for further optimization by better abstracting hardware differences and developing more complex scheduling policies.

Summary updated

Abstract

Power has emerged as a first-order design constraint in modern processors and has energized microarchitecture researchers to produce a growing number of power optimization proposals. Almost in tandem with the move toward more energy-efficient designs, architects have been increasing the number of processing elements (PEs) on a single chip and promoting the concept of running multithreaded workloads. Nevertheless, software is still lagging behind and is often unable to exploit these additional resources -- giving rise to transactional memory. Transactional memory is a promising programming abstraction that makes it easier for programmers to exploit the resources available in many- core processor systems by removing some of the complexity associated with traditional lock-based programming. This paper proposes new techniques to merge the power and transactional memory domains.
BibTeX
@misc{hughes2011optimizing,
  author = {Clay Hughes and Tao Li},
  title = {{Optimizing throughput/power trade-offs in hardware transactional memory using DVFS and intelligent scheduling}},
  year = {2011},
  doi = {10.1145/1995896.1995918},
}

Loading related content...

Figure