Abstract:
Cloud computing is now a staple of a modern software development with provisioning of scalable
and on-demand resources. Amazon Web Services (AWS), the cloud services leader, offers
great variety of types of the Elastic Compute Cloud (EC2) instances, and it makes the decision
about the adequate instance to be chosen before deployment a difficult task. A poor decision
will result in a reduction in performance, redundant costs, and slow deployment times. Available
tools like AWS Compute Optimizer and Instance Type Finder are based on CloudWatch telemetry
which means that applications must be deployed first and then it can recommend something,
which in turn adds additional cost and delays. This study suggests the pre-deployment
EC2 instance recommendation framework which is independent of cloud-generated telemetry.
The proposed system analyzes local workload behavior using a hybrid prediction approach that
combines machine learning with rule-based reasoning. System-level and application-level profiling
tools are used to collect performance metrics from representative workloads, including
CPU-intensive, memory-intensive, I/O-intensive, and mixed workloads. The collected metrics,
such as CPU usage, memory consumption, disk throughput, and network activity, are preprocessed
and transformed into structured feature vectors. In parallel, an EC2 instance specification
dataset is constructed using official AWS documentation. A supervised XGBoost classifier is
then applied to map workloads to the most suitable EC2 instance family, with initial labels generated
through rule-based feature matching. After identifying the instance family, a secondary
rule-based decision layer selects the specific instance type based on vCPU requirements, memory
demand, network performance, and EBS usage patterns. To improve transparency and user
understanding, a Retrieval-Augmented Generation (RAG) module retrieves relevant AWS documentation
to support each recommendation.
