UNISS, UPM – AI models for WL characterization
| ID: R6 | Licence: 3-Clause BSD |
Owner: UNISS UPM |
Contacts: claudio.rubattu@uniss.it mfadda1@uniss.it pruiu@uniss.it rdelussu@uniss.it |
joseandres.otero@upm.es alfonso.rodriguezm@upm.es francesca.palumbo@unica.it francesco.ratto@unica.it |
| Short Description | Approaches and methods for workload characterization (e.g. traditional statistical methods and ML-based models) to estimate different KPIs, such as execution latency or throughput and power or energy consumption, at deployment time and at run time. They rely on a measuring infrastructure that gathers execution metrics and on a dynamic modelling methodology that keeps the estimation models up to date during system operation. They are applied to the characterization of containers before deployment, based on resource usage metrics, and to the characterization of dynamically reconfigurable accelerators. | |
| Deployment time | Run time | |
|---|---|---|
| Key features | Docker containers; resource usage metrics that do not depend on the application and need no access to its source code. | Incremental modelling; non-intrusive measurement infrastructure; models can be combined across nodes with federated learning. |
| Require | Docker container support. | The MYRTUS monitoring infrastructure for FPGA-based nodes integrated within the computing nodes, to gather run-time power/performance metrics; Linux OS support to run Python. |
| Provide | Resource-related information. | Power/performance estimation of the combination of multiple accelerators on the node, and of the impact of the interference among accelerators. |
| Input | Container. | Measurements from the node and the scheduled workload. |
| Output | Files with the collected metrics and their summary statistics (CSV, JSON). | Trained models (Python), which can be consulted at run time to obtain estimation metrics. |
| User | Software developers. | Node Manager (R11) and run-time schedulers. |
| Benefits for the user | Efficient deployment of workloads. | Continuous optimization of the node operation. |
| Download | github.com/RDelussu/WL-Characterization | Monitoring infrastructure: github.com/des-cei/fpga-monitor Run-time models: please contact the developers listed above. |
| Role in MYRTUS | Characterization of the workload before deployment. | ML-based models used within the Node Manager (R11) to take run-time decisions on the optimal resource configuration. |
| TRL@M0 | 2-3 | |
| TRL@M36 | 4 | |
General description
Workload characterization describes how workloads use computing resources, to support orchestration decisions, trigger reallocations and train learning-based models. Besides the deployment-time and run-time approaches described here, resource profiles of the application components are also built for the DCOP-based Workload Manager, where they feed the simulations used to evaluate alternative allocations.
Deployment time
The workload characterization tool monitors and analyzes the execution of Docker containers under controlled resource constraints. It provides a graphical user interface to configure container parameters (CPU, memory, GPU and sampling settings) and automatically collects system and container-level metrics during execution. It supports CPU frequency monitoring, optional GPU statistics via NVIDIA utilities, and detailed logging of performance indicators over time. All collected data are stored in structured CSV and JSON files for subsequent workload characterization and performance analysis.
Run time
The run-time workload characterization methodology provides dynamic power and performance models for reconfigurable multi-accelerator nodes. A non-intrusive monitoring infrastructure collects power and performance information during workload execution, which is processed to incrementally train data-driven models at run time. The resulting models estimate CPU power consumption, programmable-logic power consumption and task execution time for different accelerator configurations, including the effects of concurrent accelerator execution. By continuously incorporating new execution observations, the characterization adapts to changes in workload behaviour and provides up-to-date information to run-time resource-management mechanisms.
MYRTUS Extension/Contribution
Deployment time
This work introduces a unified tool capable of capturing fine-grained, time-resolved metrics from containerized executions under controlled resource constraints. Beyond standard container statistics, it integrates CPU frequency monitoring, optional GPU performance tracking, and configurable execution parameters through a graphical interface, enabling a more detailed analysis of workload behaviour. The main contribution lies in improving the accuracy and reproducibility of workload characterization across heterogeneous execution scenarios, producing structured CSV and JSON outputs that support systematic performance analysis and validation. The tool has been validated with synthetic CPU-, memory-, I/O- and network-intensive workloads and with containerized video analysis applications using GPU-accelerated deep learning inference.
Run time
Within MYRTUS, a run-time workload characterization methodology has been developed to model the power and performance behaviour of dynamically reconfigurable multi-accelerator nodes during system operation. The methodology combines a non-intrusive monitoring infrastructure with incremental machine-learning models that are continuously updated from execution observations. These models estimate CPU power consumption, programmable-logic power consumption and task execution time for different accelerator configurations, including the effects of concurrent execution and interference between accelerators.
The modeling process is supervised at run time to limit unnecessary training overhead while maintaining model accuracy as workload conditions evolve. The resulting characterization can be queried by the Node Manager and scheduling mechanisms to evaluate alternative accelerator configurations and support run-time resource-management decisions. The approach has also been extended with Federated Learning, so that knowledge acquired independently by different nodes is combined into shared workload models without centralizing the underlying monitoring traces.
Technology Benefits and Innovations
Deployment time
The core innovation of this result lies in the development of a unified framework for workload characterization in containerized environments. The framework combines configurable resource-constrained execution with fine-grained monitoring of CPU, memory, network, I/O and GPU metrics, enabling a comprehensive view of workload behaviour. By integrating data collection, execution control and automated reporting within a single tool, it provides a reproducible and extensible approach for analyzing heterogeneous workloads across different computing platforms.
Current monitoring solutions often focus on isolated performance metrics or provide limited support for correlating workload behaviour with resource constraints in containerized environments. This result addresses this gap by offering synchronized collection of system-level and container-level metrics, including CPU frequency and GPU utilization data, while maintaining a lightweight and configurable execution environment. As a result, it enables more accurate workload characterization, facilitates performance analysis under controlled conditions, and supports the evaluation of diverse workload types, contributing to improved resource management and system optimization.
Run time
The core innovation lies in maintaining workload-characterization models directly during system operation rather than relying exclusively on static or design-time profiling. Power and performance observations obtained from the execution of hardware accelerators are continuously incorporated into lightweight data-driven models, allowing the characterization to evolve together with the workload. The resulting models capture the behaviour of different accelerator configurations and the impact of concurrent execution, providing information that can be directly exploited by run-time resource-management mechanisms.
Dynamic cloud-edge environments also introduce a limitation when characterization is performed independently at each node, since each model can only learn from the workloads locally observed. The federated extension addresses this limitation by enabling characterization knowledge acquired at different nodes to be combined while keeping the original monitoring data local. This provides a path towards a broader and more representative characterization of workloads across the continuum, while retaining the ability of individual nodes to adapt their models at run time.
Users
Deployment time
The expected users of this result include researchers and academics working on workload characterization, performance analysis, cloud computing and resource management. The tool is also relevant for system administrators, performance engineers and DevOps practitioners who need to monitor and evaluate the behaviour of containerized applications under different resource constraints. In addition, developers and practitioners in domains such as High-Performance Computing (HPC), artificial intelligence and data-intensive computing may use the tool to analyze workload characteristics, benchmark applications, and support optimization and deployment decisions in heterogeneous computing environments.
Run time
The main users of the run-time characterization result are run-time resource-management and scheduling components operating on heterogeneous and reconfigurable computing nodes, including the MYRTUS Node Manager and higher-level orchestration mechanisms. The technology is also relevant for developers and researchers working on adaptive FPGA-based systems, heterogeneous edge computing and run-time optimization who require continuously updated estimates of workload power consumption and performance.
Application Domains
Deployment time
The proposed tool can be applied across a wide range of domains where understanding workload behaviour is critical. These include cloud and edge computing environments, where resource allocation and performance optimization are essential, as well as High-Performance Computing (HPC), where detailed analysis of CPU and GPU utilization is required. It is also suitable for containerized microservices architectures, enabling performance monitoring and benchmarking under realistic deployment conditions. Furthermore, it can support research and industrial scenarios involving data-intensive, AI and machine learning workloads, as well as systems engineering tasks focused on resource management, system profiling and performance evaluation across heterogeneous computing platforms.
Run time
The methodology targets heterogeneous and reconfigurable computing systems in which workload behaviour and resource availability may change during operation. Relevant application domains include edge and cloud-edge continuum computing, FPGA-based acceleration, adaptive embedded systems and distributed cyber-physical systems. Within MYRTUS, the generated workload models provide run-time knowledge that can be used by node-level scheduling and resource-management mechanisms to evaluate alternative accelerator configurations and support the efficient execution of dynamically arriving workloads. The federated extension further enables characterization knowledge to be shared across distributed nodes, providing a basis for future continuum-level scheduling, offloading and workload-placement decisions.
References
- J. Encinas et al., “Leveraging Incremental Machine Learning for Reconfigurable Systems Modeling under Dynamic Workloads”, ACM Transactions on Reconfigurable Technology and Systems, 18(1), 2025.
- E. Fernandez, J. Encinas, A. Otero, “Modelling of Hardware Accelerators in the Cloud-Edge Continuum with Federated Learning Techniques”, Zenodo, 2026.
- Deployment-time characterization tool: github.com/RDelussu/WL-Characterization
