Work

Efficient Directed Acyclic Graph Application Deployment Strategies in Single-tier and Multi-tier Computing Environments

Year: 2024
Type: dissertation
Abstract: In today’s fast-evolving tech landscape, data-driven applications, especially those involving DeepNeural Network (DNN) models and Directed Acyclic Graphs (DAGs), are increasingly prominent. These applications typically require refined, pre-processed data, primarily handled at the edge tier. However, the computation-heavy tasks of DNN models and DAG workflows necessitate substantial computational power, leading to their deployment in multi-tier environments. In single-tier placement, we address the joint optimization of two conflicting objectives: energy consumption at fog nodes and communication costs of applications. We treat this as a game between fog nodes and applications, where each agent is modeled to control one specific objective. Initially, we developed an Iterative Combinatorial auction-based distributed algorithm, which operates independently of any central controller or auctioneer. Subsequently, we designed an autonomic algorithm that uses a Markov approximation method. This algorithm is capable of adjusting placements to accommodate changes in the edge infrastructure and network variations. We extend Kubernetes’ scheduling capabilities to accommodate large, containerized applications in fog environments. By modifying its default scheduler to handle split containers, our approach minimizes network communication while maintaining functionality, as demonstrated through real-world IoT deployments. In a multi-tier environment, we consider the efficient placement of DNNs and other DAG-based workflows. Our research uniquely identifies infrastructure and bandwidth-independent split points within various DNN models that consistently lead to efficient latency or cost partitioning. This novel insight significantly simplifies the computational complexity of DNN partitioning, making our DNNSplit strategy suitable for models with numerous layers. For DAG applications, we advocate a top-down approach that prioritizes latency or cost optimization in multi-tier environments to meet QoS and SLA requirements. Our wDAGSplit strategy efficiently resolves bandwidth contention and enables parallel executions, substantially reducing completion times.
Language: English
Open Access status: green
OpenAlex ID: W7133005651
Cites: 106
Cited by:
Related to: 10