Machine Learning based resource allocation for cloud applications

The rapid cloud migration of the enterprises coupled with Ansys’ growing emphasis on cloud computing (thanks to PyANSYS!) gives a noticeably clear indication of where the future lies. While Ansys SaaS (Software as a Service) solutions have clear advantages compared to on-premises installations, they may still cost a fortune especially for long and memory intensive analysis. The user is often stuck in this dilemma of the number of workers to launch. If it launches a lot of workers, there is a chance that the application is not able to utilize them because of the limitation in order of parallelism that it offers resulting in wastage of resources. Similarly, a run with fewer workers may consume a lot of run time and may crash eventually due to high peak memory. The proposed solution will help improve the user experience by predicting the resources needed to complete the analysis. It may also help reduce the cloud billing cost and/or reduce the overall runtime to complete an analysis. The machine learning model used will be based on the application that is launched and will be trained with the help of domain experts. The first prototype will be built using Ansys Redhawk-SC. Redhawk-SC does EMIR analysis (which may take up to 24 hours+ on production designs), there is a series of interdependent views which are created, each offering different degrees of parallelism. In the current flow, the workers are reserved before the start of the analysis, and they sit idle, in case the view running does not offer a high degree of parallelism. The idea is to evaluate the resources needed at the start of each view to optimally utilize them and thus reduce the billing cost. There will be a separate model for each view which will depend on the attributes of the input file for that view (like file size/number of instances) and the attributes from preceding dependent views. Once done successfully, it can be enhanced to other applications as well.