KCL, HIRO – AI-enabled network manager
| ID: R12 | Licence: To be defined |
Owner: KCL HIRO |
Contacts: yansha.deng@kcl.ac.uk na.2.yan@kcl.ac.uk |
| Short Description | A DRL-based bandwidth allocation network manager to enable latency-aware resource orchestration across cloud, fog, and edge nodes. |
| Key features | – Uplink/downlink bandwidth allocation per flow (uplink for edge-to-fog flows, downlink for fog-to-edge flows). – DRL policy trained to improve the flow success rate and reduce latency violations. – Estimation of the transmission latency of each flow and check against its latency requirement. – REST interfaces for asynchronous allocation requests, status and results. |
| Require | Python 3.9, PyTorch, DRL libraries (e.g. PPO), JSON input interfaces, simulated network traces. |
| Provide | Source code, DRL models. |
| Input | A batch of flow requests (source and destination nodes, data size, communication direction, latency requirement) and the current network state (available uplink and downlink bandwidth). |
| Output | Bandwidth allocation decisions per flow; estimated transmission delay; success or violation of the latency requirement for each flow; reward logs. |
| User | Workload Manager, Deployer. |
| Benefits for the user | Enables dynamic, bandwidth-aware workload placement; increases the flow success rate while maintaining latency constraints. |
| Download | Please contact the developers listed above. |
| Role in MYRTUS | Provides the decision-making logic for the Network Manager (NeM) of the MIRTO cognitive engine. |
| TRL@M0 | 1 |
| TRL@M36 | 3 |
General description
The Network Manager dynamically manages network resources across the edge, fog and cloud layers, to support the execution of service chains on distributed and heterogeneous computing resources. When the Workload Manager places software components on different compute nodes, their communication flows compete for the available bandwidth: static network management cannot cope with rapidly changing workloads and connectivity demands.
For each request, the Network Manager receives the communication flows generated by the Workload Manager, each with its source and destination, data size, direction and latency requirement. It retrieves the available uplink and downlink bandwidth, allocates it among the flows using pre-trained deep reinforcement learning models, estimates the transmission latency of each flow and checks whether its latency requirement is satisfied. The result is returned to the Workload Manager, so that placement decisions take communication into account.
MYRTUS Extension/Contribution
Within MYRTUS, the Network Manager has been implemented as a component of the MIRTO cognitive engine, with OpenAPI-based interfaces to submit bandwidth-allocation requests and to retrieve their status and results asynchronously. The DRL models are trained offline with simulated flow requests and network conditions, using Proximal Policy Optimization (PPO); since the number of flows determines the size of the problem, a separate model is trained for each supported number of flows, and the Network Manager selects the right one for each request.
The DRL algorithms of the Network Manager belong to the same family as those of the DRL-based Workload Manager (result R9), with which it has been integrated: for each service offloaded from the edge to the fog, the Workload Manager submits the corresponding uplink and downlink flows and retrieves the allocated bandwidth and the estimated latency. In the integrated example, all flows met their latency requirements.
Technology Benefits and Innovations
The Network Manager replaces static bandwidth allocation with a learning-based policy that adapts to the flows actually requested and to the bandwidth actually available. Because it works on requests from the Workload Manager and returns per-flow latency estimates, network conditions become part of the placement decision rather than a constraint discovered after deployment.
Users
The expected users are developers and integrators of orchestration and resource-management platforms for the edge-fog-cloud continuum, who need to place distributed applications taking communication into account, especially where bandwidth between layers is limited and services have latency requirements. The result is also relevant for researchers working on learning-based management of network resources in distributed systems.
