MLFV: uma abordagem consciente do estado da rede para orquestração de cadeias de funções de aprendizado de máquina

Abstract

Machine Learning as a Service (MLaaS) platforms are allowing the application of Machine Learning (ML) techniques from anywhere, and at any time. These platforms, in general, are hosted in the cloud and have a scalable infrastructure with high processing power; however, they have some disadvantages, such as the need to send data to the cloud. ML on the Edge (Edge Computing) is emerging as an option to tackle some limitations imposed by these platforms, reducing the latency and bandwidth usage; furthermore, it avoids data privacy and security issues by keeping the data on the local network. However, the application of ML on the edge still presents research challenges, such as the orchestration of ML functions considering the network state and the computational capabilities of the nodes. In this sense, network-aware orchestration services, provided by Network Function Virtualization (NFV) platforms can be a promising approach to manage the ML tasks placement. This work proposes the MLFV (Machine Learning Function Virtualization), a network-aware approach that explores the NFV environment to orchestrate the execution of ML function chains; these chains represent the execution flow of the ML functions that can be grouped as sequential and/or parallel activities. The MLFV implements a model for placing chains of ML, considering constraints on CPU, memory, required libraries and the network overload aiming to reduce the overall execution time of all functions in a chain. This model distributes the functions in order to reduce the network overload, the execution time, especially in cases where the network presents some instability. To evaluate the MLFV proposal a case study in the geotechnical area was conducted, using soil data to reproduce the soil classification process through two ML function chains; these chains, implemented by MLFV, were created based on the Knowledge Discovery in Databases process (KDD). The results showed that MLFV achieved, on average, a 25% reduction in the execution time compared to cloud (MLaaS) and edge approaches in a stable network connection scenario. When some computational nodes had bandwidth constraints, MLFV was able to identify these limitations, allocating the ML tasks on hosts with stable connections. The other approaches were unable to detect these instabilities, resulting in a 400% increase in the overall chain execution time.