TL;DR
This paper proposes a novel approach to map and partition OpenCL applications onto heterogeneous CPU-GPU MPSoCs by considering the thermal behavior of the system. By profiling temperature alongside performance and energy, they achieve up to a 13% reduction in average chip temperature, leading to improved long-term reliability without sacrificing performance.
What Problem Does This Paper Address?
Modern embedded systems increasingly rely on heterogeneous Multi-Processor Systems-on-Chips (MPSoCs) that combine different types of processing cores, such as CPUs and GPUs. This heterogeneity offers significant potential for performance and energy efficiency by leveraging the strengths of each core type. However, effectively utilizing these diverse cores for applications, especially those written in parallel programming models like OpenCL, is challenging. Existing approaches often focus on performance and energy consumption for mapping and partitioning tasks. A critical, yet often overlooked, aspect is the thermal behavior of the system. High temperatures can degrade performance over time and significantly reduce the lifespan of the chip. This paper addresses the gap in existing research by incorporating temperature-aware considerations into the mapping and partitioning process for OpenCL applications on CPU-GPU MPSoCs, aiming to improve reliability.
Methods
The authors propose a temperature-aware mapping and partitioning methodology. The core of their approach involves:
- Generating Mappings: For a given application, they explore various combinations of CPU (big and LITTLE cores) and GPU core assignments. The total number of possible mappings is calculated based on the available CPU cores and the single GPU.
- Evaluating Mappings at Various Partitions: For each generated mapping, they systematically vary the fraction of application workload (threads) executed on the CPU versus the GPU. During these executions, they measure:
- Execution Time (ET): The time taken for the application to complete.
- Energy Consumption (EC): Measured using on-board power sensors.
- Temperature Profile (TEMP): Captured by on-board temperature sensors for both CPU and GPU cores. The average temperature over the execution is used as it directly impacts reliability.
- Finding Performance and Energy Optimized Set: From the evaluated design points, they identify a subset that meets the application's performance requirements and exhibits low energy consumption.
- Selecting the Temperature and Energy Optimized Design Point: From this optimized set, they select the specific mapping and partition that minimizes the product of energy consumption and average temperature. This aims to find a balance that offers good performance, low energy, and a favorable thermal profile.
The key novelty lies in explicitly integrating temperature profiling into the decision-making process for mapping and partitioning, moving beyond purely performance and energy optimization.
Datasets
The evaluation uses applications from the Polybench benchmark suite [12]. Specifically, the paper mentions using the GPU versions of these benchmarks, which are adapted to run on both CPU and GPU cores. The specific applications used include: SYR2K, SYRK, CORRELATION, COVARIANCE, 2MM, 2DCONV, GEMM, and MVT. The experiments were conducted on an Odroid-XU3 platform with a modified Ubuntu Linux Kernel 3.10.96, equipped with a heat sink and fan.
Key Insights and Ideas
- Temperature is a critical factor for reliability: The paper strongly emphasizes that high temperatures significantly impact the long-term reliability and lifespan of MPSoCs, a factor often overlooked in traditional performance/energy optimization.
- Trade-offs exist between performance, energy, and temperature: Different partitions of workload between CPU and GPU lead to varying execution times, energy consumption, and thermal profiles. This creates a design exploration space where optimal choices can be made.
- Application-specific behavior: The optimal mapping and partitioning strategy is not universal; it varies significantly between different applications, highlighting the need for dynamic or adaptive approaches.
- Temperature profiling reveals hidden optimization opportunities: By actively measuring and considering temperature, the authors can identify configurations that might be slightly less performant or energy-efficient in the short term but offer substantial long-term benefits due to lower thermal stress.
Key Results
The proposed temperature-aware mapping and partitioning approach demonstrated significant improvements:
- Reduced Average Temperature: The method achieved up to a 13% saving in average chip temperature compared to a state-of-the-art approach that only considered performance and energy.
- Performance Maintenance: These temperature savings were achieved without sacrificing performance requirements. In some cases, there was a slight overhead in performance (e.g., 8.88% for application 2D), but this was traded for substantial temperature reductions (e.g., 10.73% for application 2D).
- Significant Temperature Savings: For specific applications like SYRK, the temperature savings reached up to 13.83%, translating to an average temperature reduction of around 8°C.
Limitations and Future Work
- Static Partitioning: The current approach performs static partitioning before runtime. The authors suggest that extending this to run-time partitioning would allow for more dynamic adaptation to changing workloads and system conditions.
- Single Application Focus: The current evaluation focuses on individual applications. Future work could explore how to extend the methodology to handle multiple applications executing concurrently on the MPSoC.
- DVFS Extension: The authors propose extending the method to incorporate Dynamic Voltage and Frequency Scaling (DVFS) for each computing module (A15, A7, and GPU cores) separately, which could further optimize energy and thermal management.
- Experimental Platform: While the Odroid-XU3 is a representative MPSoC, testing on a wider range of hardware platforms could validate the generalizability of the approach.
Summary updated