INCREASING THE EFFICIENCY OF AI TASK DEPLOYMENT FOR CLOUD-EDGE ENVIRONMENTS

Abstract

Relevance of the research topic. Contemporary literature on AI task deployment states the importance of MLOps principles and the ways of implementing them without optimizing the individual steps. Cloud-Edge hybrid architecture requires specific measures to be taken in order for these principles to be applicable to IoT devices, which have low resources. Introducing the same software to such systems may cause increased faults and decrease uptime. To modify the pipeline, a suite of techniques is introduced. Measurements such as time and disk size utilization are made to quantify the impact on the system.

Target setting. Different articles show variety in CI/CD process implementation and require analysis and modification of its steps in order to increase the efficiency and usability on different platforms. To fully utilize the capabilities of IoT systems, modifications are required that will make the images more lightweight and the pipeline faster.

Actual scientific researches and issues analysis. Current scientific research shows that CI/CD is widely used and researched in order to be applied to AI problems. However, there is much less literature on how to optimize the CI/CD pipelines and its components. These articles propose different optimization techniques that apply to different areas of research show promise and can be usable in AI systems.

Uninvestigated parts of general matters defining. Contemporary solutions provide minimal insight into the performance of such systems in comparison with other approaches. MLOps principles have not been put into testing in end-to-end pipelines to quantify their performance and outline areas of improvement. General improvement of each step is required to launch the pipeline efficiently on IoT/Edge hardware and to fully automate it using Cloud.

The research objective. The goal of the article is to investigate methods of improving the deployment task of MLOps pipelines by increasing the uptime and decreasing build times.

Presentation of the main material. The foundation of this approach is to create an efficient CI/CD pipeline. Continuous integration (CI) is a process that integrates code changes into the existing code base to resolve conflicts between changes made by different developers. It includes build and test phases that involve automatic code generation and testing. The planning and coding steps are tied to CI, making deployment more efficient.
Container optimization in this context includes a number of measures aimed at reducing the size and resource usage of the container. The purpose of optimization is to accelerate the construction and deployment speed of container applications at the network edge, as well as to reduce the use of system resources.

Conclusion. In this paper, we proposed a methodogical approach to improvement of AI-systems. MLOps practices ensure the effectiveness and quality of artificial intelligence solutions over time. The core of this approach is creating an effective CI/CD pipeline, which includes continuous integration (CI), planning and coding phases, and Continuous Delivery (CD), which keeps applications ready for production deployment.

Container optimization is a method that considers individual system elements and their impact on its functioning. It aims to speed up the build and deployment of containerized applications at the edge of the network and reduce system resource use. This can be achieved through configuration optimization, hard and soft restrictions, using a reduced container image, monitoring and adjusting container resources, and using lightweight base images. These measures help maintain system resiliency, improve system scaling decisions, and remove redundant resources.

Another approach is to use sequential downloads. Literature shows that parallel downloads are usually prone to blocking whereas sequential downloads allow for speedier image builds. Along with other techniques, like dependency management, as shown in the table, these measures substantially lowered the amount of disk usage for images.

Uninvestigated parts. To further improve the proposed methods, a further look can be taken at AI model optimization. Using technqiues like pruning, AI model can be fitted to a smaller device, like Raspberry PI nano. For a more robust testing another metric of resource utilization can be utilized, especially for specific optimization techniques. These approaches contribute to refining the performance and applicability of AI models across various platforms and scenarios.

Another area of research may be utilizing different devices and making load test devices. This research can provide additional information by assessing how various hardware configurations and environments influence the performance and scalability of systems or applications under different loads. By making experiments across a range of devices, we can gain insights into optimizing performance, resource allocation, and overall system efficiency across varied technological contexts. This exploration contributes to advancing the understanding and implementation of robust, adaptable systems capable of meeting diverse operational demands.