Monitoring and Observability

Partners:
UPM, UNICA, UNISS (FPGA monitoring)
ABI, HIRO, CRF (monitoring of their nodes)

Role in MYRTUS:
Visibility of the infrastructure and run-time information for orchestration

Description

Monitoring and observability serve two purposes in the MYRTUS Reference Infrastructure: they give visibility of the state of the heterogeneous resources spread across the edge, fog and cloud layers, and they provide the run-time information that the MIRTO cognitive engine needs to manage workloads and resources. Monitoring data is stored in the Distributed Knowledge Base.

MYRTUS combines two complementary approaches. Established tools, mainly Prometheus and Grafana, collect and display system- and cluster-level metrics. A dedicated monitoring infrastructure, developed in MYRTUS for FPGA-based nodes, captures synchronised power and accelerator performance traces while workloads run.

Fog

SenseIQPrometheus · GrafanaSystem, container and network metrics, also collected from a remote SenseIQ instance connected through Liqo.

Fog Micro Data CenterPrometheus · Grafana · KeplerCluster usage metrics, extended with node-level energy data (CPU, memory and package power).

Edge

Traffic camerasPrometheus · MYRTUS node monitorSystem and pod metrics shown in Grafana; camera metrics pushed to the DKB.

FPGA-based nodesMYRTUS FPGA monitoring frameworkSynchronised power and performance traces of the hardware accelerators.

How the nodes of the MYRTUS infrastructure are monitored.

Monitoring infrastructure for FPGA-based nodes

On FPGA-based nodes, system-level monitoring shows the overall state of the node, but not what happens inside the hardware accelerators. MYRTUS therefore developed a monitoring infrastructure that records power consumption and accelerator performance at the same time, so that each execution event can be matched with the power behaviour it causes. It is used on the edge nodes based on HMPSoC accelerators and on the RISC-V-based processor.

The framework is modular and composable: its parts are combined according to the node, so that very different nodes share a common architecture and interface. The monitoring logic is a configurable hardware block placed in the FPGA design next to the accelerators, with drivers and libraries to configure it and read the traces.

PowerFrom the sensors on the FPGA board or, for higher sampling rates and more detail, from an external measurement board; whole board or single power rails.

PerformanceFrom the accelerators’ start and stop signals, from their bus transactions, or from other internal signals chosen by the user.

PlatformEmbedded, soft-core or host processors; different FPGA platforms; Linux-based and bare-metal systems.

Synchronised tracespower and performance, timestamped on a common time reference

Offline analysisTraces stored as files.

Run-time useTraces kept in local buffers.

VisualisationTraces sent to visualisation tools.

The modular monitoring framework for FPGA-based nodes.

Within MYRTUS, the traces characterise how FPGA-based workloads behave at the edge and provide telemetry to the resource management of MIRTO.

References

J. Encinas et al., “Modular and Composable Framework for Multipurpose Monitoring in Heterogeneous FPGA-Based Systems”, in Proc. SAMOS 2025. Zenodo
Prometheus, Grafana, Kepler

Monitoring and Observability