Towards a Methodology for Creating Time-critical, Cloud-based CUDA Applications

CUDA has been used in many different application

domains, not all of which are specifically image processingrelated.

There is the opportunity to use multiple and/or distributed

CUDA resources in cloud facilities such as Amazon

Web Services (AWS), in order to obtain enhanced processing

power and to satisfy time-critical requirements which cannot be

satisfied using a single CUDA resource. In particular, this would

provide enhanced ability for processing Big Data, especially in

conjunction with distributed file systems (for example). In this

paper, we present a survey of time-critical CUDA applications,

identifying requirements and concepts that they tend to have

in common. In particular, we investigate the terminology used

for Quality of Service metrics, and present a taxonomy which

summarises the underlying concepts and maps these terms to the

diverse terminology used. We also survey typical requirements

for developing, deploying and managing such applications. Given

these requirements, we consider how the SWITCH platform can

in principle support the entire life-cycle of time-critical CUDA

application development and cloud deployment, and identify

specific extensions which would be needed in order fully to

support this particular class of time-critical cloud applications.