PRODUCTIVITY ESTIMATION OF SERVERLESS COMPUTING

Cloud computing has enabled organizations to focus less on their IT infrastructure and more on their core products and services. In fact, Cloud is no longer viewed as an alternative to hosting infrastructure. Serverless computing is a technology, also known as function-as-a-service, that gives the cloud provider complete management over the container function run on as necessary to serve requests. As a result, the architectures remove the need for continuously running systems and serve as event driven computing.Serverless computing presents new opportunities to architects and developers of Cloud-oriented solutions. Primarily, it provides a simplified programming model for distributed Cloud-based systems development, with the infrastructure abstracted away. It is no longer the concern of the developer to manage load balancers, provisioning and resource allocation (although system im-plementers need to be aware of such things). This reduced focus on operational concerns should allow greater attention to be paid to delivering value, functionality and an ability to adapt rapidly to changes. Such issues as deployment, monitoring, quality of service and fault tolerance are moved into the hands of the Cloud provider and still need to be actively considered and managed. Serverless computing is still in its infancy and while the model matures further, tools will be created to allow developers and architects to create patterns and processes to fully exploit the advantages of the Serverless model.This paper explores the performance profile of a Serverless ecosystem under low latency and high availability. The results of application and performance tests for image recognition by using neural networks are presented. The proposed implementation uses open source libraries and tools: TensorFlow for the study of machine learning and LabelImg for data preparation. A correlation between the amount of experimental training data and recognition accuracy is studied and shown. For experiments, the software package was developed using the Python scripting programming language and .Net technology. The developed software showed excellent accuracy of recognition using regular computer with low-cost hardware. Interaction of the client side with the “server” is carried out using HTTP-requests in any browser with low-speed network connection.