A Highly Accurate Energy Model for Task Execution on Heterogeneous Compute Nodes
| Autor: | Marco Platzner, Achim Losch |
|---|---|
| Rok vydání: | 2018 |
| Předmět: |
020203 distributed computing
Optimization problem business.industry Computer science Distributed computing Control reconfiguration Symmetric multiprocessor system 02 engineering and technology Energy consumption 020202 computer hardware & architecture Scheduling (computing) 0202 electrical engineering electronic engineering information engineering Task analysis Electricity business Field-programmable gate array Computer Science::Operating Systems |
| Zdroj: | ASAP |
| DOI: | 10.1109/asap.2018.8445098 |
| Popis: | Heterogeneous computing with CPUs, GPUs, and FPGAs has strongly gained interest in the last years. While scheduling and optimization problems for runtime have been widely studied, optimizing for energy-related metrics has become an emerging topic only recently due to rising electricity costs and the difficulties of thermal management. Energy-optimizing schedulers need to predict the effect of single task-resource assignment decisions on the consumed energy as well as the energy consumptions for complete schedules. In this paper, we present a highly accurate energy model for heterogeneous compute nodes. Compared to previous work, that differentiated between static and dynamic energy consumption and included idle energies, the new and accurate model looks into the impact of host-side activities of tasks executed on accelerators, covers more performance and power states of the devices, and considers resource-specific features such as reconfiguration processes for FPGA tasks. We present experiments and analyses of task executions on CPU, GPU, and FPGA, which allowed us to derive a set of critical model refinements over previous work. For evaluation, we compare the predictions of our model with previous work and with real measurements gained during the execution of 200 randomly generated schedules. The results show the high accuracy of our new energy model, with prediction errors as low as 0.3% and 0.4%, depending on the task set characteristic. |
| Databáze: | OpenAIRE |
| Externí odkaz: |
|