研究目的
Investigating the development and implementation of a power-aware scheduling algorithm for heterogeneous CPU-GPU architectures to reduce peak power consumption while maintaining computational efficiency.
研究成果
The power-aware scheduling algorithm effectively reduces peak power consumption by up to 10% compared to systems without any power-aware policy, with a negligible impact on performance. It also mitigates the worst case power scenario by up to 24%, demonstrating its potential to reduce system and service costs without significantly affecting computational efficiency.
研究不足
The experimental results are dependent on the target architecture used in the study. A new characterization is needed for different architectures or when new kernels are added to the library.
1：Experimental Design and Method Selection:
The study involves the development of a power-aware scheduling algorithm for heterogeneous CPU-GPU architectures, focusing on workload distribution to reduce peak power.
2：Sample Selection and Data Sources:
The algorithm is tested on a real CPU-GPU heterogeneous system with workloads generated using a Markov chain model.
3：List of Experimental Equipment and Materials:
The setup includes computing nodes with Intel Xeon E5520 CPUs, NVIDIA GeForce GTX 590 GPUs, and a power measuring system composed of Hall effect current sensors and a microcontroller.
4：Experimental Procedures and Operational Workflow:
The algorithm is evaluated by executing and measuring 10 workloads of 1,000 job requests, comparing performance with and without the power-aware characteristic.
5：Data Analysis Methods:
Performance is evaluated based on peak power reduction, increase in computation time, and energy consumption, using statistical techniques and software tools for analysis.
独家科研数据包，助您复现前沿成果，加速创新突破
获取完整内容
