Item type |
Symposium(1) |
公開日 |
2022-12-20 |
タイトル |
|
|
タイトル |
Job Pipeline Mechanism of GPU Node for Multi-node Edge Computing (MEC) Systems |
タイトル |
|
|
言語 |
en |
|
タイトル |
Job Pipeline Mechanism of GPU Node for Multi-node Edge Computing (MEC) Systems |
言語 |
|
|
言語 |
eng |
資源タイプ |
|
|
資源タイプ識別子 |
http://purl.org/coar/resource_type/c_5794 |
|
資源タイプ |
conference paper |
著者所属 |
|
|
|
Shibaura Institute of Technology |
著者所属 |
|
|
|
Shibaura Institute of Technology |
著者所属 |
|
|
|
Shibaura Institute of Technology |
著者所属(英) |
|
|
|
en |
|
|
Shibaura Institute of Technology |
著者所属(英) |
|
|
|
en |
|
|
Shibaura Institute of Technology |
著者所属(英) |
|
|
|
en |
|
|
Shibaura Institute of Technology |
著者名 |
Taiki, Miyakawa
Yanzhi, Li
Midori, Sugaya
|
著者名(英) |
Taiki, Miyakawa
Yanzhi, Li
Midori, Sugaya
|
論文抄録 |
|
|
内容記述タイプ |
Other |
|
内容記述 |
A Multi-node Edge Computing (MEC) system that integrates hardware with different accelerators, such as GPUs and FPGAs, has been proposed for high-performance computing applications with low power consumption. Generally, a GPU node that connects to the MEC system provides a huge computation resource for AI applications. However, the current implementation of the MEC system executes jobs serially, which causes GPU utilization to decrease when processing multiple jobs. Moreover, there is a lack of a mutual-execution mechanism that avoids resource competition on the GPU, which causes abnormal program termination and may result in incorrect processing. To improve the efficiency and reliable execution of the multiple jobs on the GPU node, we propose a job pipeline mechanism that receives multiple jobs from the requester and executes them as node programs parallelly cooperate with the CPU. Also, we propose a mechanism that provides mutual exclusion to avoid abnormal termination of the jobs. In the evaluation, we found the efficient use of the GPUs of the system and avoid abnormal termination despite the multiple job assignment. |
論文抄録(英) |
|
|
内容記述タイプ |
Other |
|
内容記述 |
A Multi-node Edge Computing (MEC) system that integrates hardware with different accelerators, such as GPUs and FPGAs, has been proposed for high-performance computing applications with low power consumption. Generally, a GPU node that connects to the MEC system provides a huge computation resource for AI applications. However, the current implementation of the MEC system executes jobs serially, which causes GPU utilization to decrease when processing multiple jobs. Moreover, there is a lack of a mutual-execution mechanism that avoids resource competition on the GPU, which causes abnormal program termination and may result in incorrect processing. To improve the efficiency and reliable execution of the multiple jobs on the GPU node, we propose a job pipeline mechanism that receives multiple jobs from the requester and executes them as node programs parallelly cooperate with the CPU. Also, we propose a mechanism that provides mutual exclusion to avoid abnormal termination of the jobs. In the evaluation, we found the efficient use of the GPUs of the system and avoid abnormal termination despite the multiple job assignment. |
書誌情報 |
Proceedings of Asia Pacific Conference on Robot IoT System Development and Platform
巻 2022,
p. 56-59,
発行日 2022-12-20
|
出版者 |
|
|
言語 |
ja |
|
出版者 |
情報処理学会 |