Distributed deep learning, studied at system scale.
DDLSim-Lab provides a reproducible environment for studying distributed deep learning across heterogeneous infrastructure, network conditions, scheduling strategies, and failure scenarios.
A research environment built for distributed systems.
DDLSim-Lab is an open-source research project focused on understanding how distributed deep learning systems behave under realistic infrastructure constraints.
The platform is designed to support experimentation with large-scale, heterogeneous, and failure-prone environments.
Researchers can investigate algorithms, scheduling policies, communication behavior, resource allocation and fault-tolerance techniques.
Systems-level experimentation
Study the interaction between distributed workloads and the infrastructure beneath them.
What the lab can investigate
The environment brings several system dimensions together so experiments can examine their interaction.
Distributed Training
Distributed deep learning workloads across multiple compute nodes.
Network Behavior
Latency, jitter, packet loss, bandwidth and communication behavior.
Scheduling
Placement, resource allocation and adaptive scheduling strategies.
Performance
Scaling behavior and system performance under changing conditions.
Fault Tolerance
Failure scenarios, recovery and resilient distributed behavior.
Edge–Cloud Systems
Workloads spanning edge resources and cloud infrastructure.
From experiment design to observation.
A simple research flow keeps experiments understandable and reproducible.
Define workload
Specify the distributed workload, resources and experimental conditions.
Configure system
Configure nodes, networking, scheduling and infrastructure behavior.
Run experiment
Execute the workload while observing system and network behavior.
Analyze results
Compare measurements and evaluate the behavior of the system.
A systems view of the experiment.
The architecture connects workload, orchestration, network behavior, compute nodes and experiment telemetry.
Bare-metal experimentation
When infrastructure is part of the experiment.
DDLSim-Lab is designed to support bare-metal deployments when experiments require direct access to networking, compute and kernel-level behavior.
- Reduced virtualization overhead for performance-sensitive measurements.
- Direct access to network interfaces for realistic communication experiments.
- Support for technologies such as SR-IOV and RDMA where required.
Built to be inspected, reproduced and extended.
DDLSim-Lab is developed as an open-source research project. The implementation can evolve with new experiments, research questions and contributions.