Towards extending the SWITCH platform for time-critical, cloud-based CUDA applications: Job scheduling parameters influencing performance

作     者：Knight, Louise Stefanic, Polona Cigale, Matej Jones, Andrew C. Taylor, Ian 

作者机构：Cardiff Univ Sch Comp Sci & Informat Queens Bldg5 Parade Cardiff CF24 3AA S Glam Wales 

出 版 物：《FUTURE GENERATION COMPUTER SYSTEMS-THE INTERNATIONAL JOURNAL OF ESCIENCE》 (下代计算机系统)

年 卷 期：2019年第100卷

页      面：542-556页

核心收录：

学科分类：08[工学] 0812[工学-计算机科学与技术（可授工学、理学学位）] 

基　　金：European Union 

主　　题：Time-critical applications CUDA Distributed cloud computing 

摘      要：SWITCH (Software Workbench for Interactive, Time Critical and Highly self-adaptive cloud applications) allows for the development and deployment of real-time applications in the cloud, but it does not yet support instances backed by Graphics Processing Units (GPUs). Wanting to explore how SWITCH might support CUDA (a GPU architecture) in the future, we have undertaken a review of time-critical CUDA applications, discovering that run-time requirements (which we call wall time ) are in many cases regarded as the most important. We have performed experiments to investigate which parameters have the greatest impact on wall time when running multiple Amazon Web Services GPU-backed instances. Although a maximum of 8 single-GPU instances can be launched in a single Amazon Region, launching just 2 instances rather than 1 gives a 42% decrease in wall time. Also, instances are often wasted doing nothing, and there is a moderately-strong relationship between how problems are distributed across instances and wall time. These findings can be used to enhance the SWITCH provision for specifying Non-Functional Requirements (NFRs);in the future, GPU-backed instances could be supported. These findings can also be used more generally, to optimise the balance between the computational resources needed and the resulting wall time to obtain results. (C) 2019 Elsevier B.V. All rights reserved.