A software platform for data center control and optimization that combines energy-aware workload scheduling, autonomous power management, and a thermo-fluid dynamic Digital Twin. Reduce consumption, improve PUE, and prevent thermal hotspots, while maintaining SLA compliance — with results measured on real production-class hardware.
| Metric | Result | How |
|---|---|---|
| Energy per unit of work | −31% | measured at the wall, replicated (p < 0.05) |
| Throughput | +44% | measured, same jobs and hardware |
| Facility energy during idle periods | up to −52% | measured, autonomous power-off on a diurnal trace |
| Annual electricity bill (typical cluster) | ~−41% | modeled from measured power states, conservative assumptions |
DCIMate's scheduling and power-management stack runs autonomously on a production-class Kubernetes cluster of eight servers spanning four hardware generations (Intel Xeon and AMD EPYC, 472 physical cores). Every energy figure below is measured at the wall with facility-grade BMC/IPMI power metering — not simulated, and not estimated from CPU counters.
Identical jobs cost very different amounts of energy depending on which server runs them — on our cluster, up to 2× for compute work and 6× for memory-bound work. Stock schedulers are blind to this: they place work wherever CPUs are free. DCIMate routes each job to the node that completes it for the fewest joules.
Draining the same fixed backlog of AI-inference jobs through stock Kubernetes and through DCIMate, on the same hardware, placement delivered 31% less energy per job and 44% higher throughput — replicated over five paired runs, winning every time (p < 0.05). Average power stayed flat: the cluster simply finished the work sooner. No new hardware, no performance trade-off.
A modern server drains roughly half its peak power while doing nothing — 200–300 W per node on our cluster — and CPU frequency tuning recovers almost none of it (we measured ~0% at the wall). The only way to reclaim those watts is to power the machine off. DCIMate consolidates work onto efficient nodes and autonomously powers off the machines it frees, waking them within minutes when demand returns. On a four-hour realistic demand cycle this cut measured facility energy by 52%.
The hard part — and the part DCIMate is built around — is knowing when powering off is safe. Done naively, it backfires: we measured that reacting to every short dip in demand makes the cluster worse on both energy and service quality, cycling hardware and starving capacity. DCIMate's hysteretic control tells a genuine lull from a brief dip: under load it performs zero power-offs, and it folds the cluster only when idleness is sustained — so it is never worse than doing nothing.
Feeding the measured power states into an annual cost model (a modeled projection, clearly distinct from the measurements above): on a typical, partly-idle cluster the two levers together cut the electricity bill by roughly 41%, about EUR 257 per server per year at EUR 0.25/kWh — scaling nearly linearly to fleets of thousands of machines, and growing larger in sites with high cooling overhead, because a powered-off server also stops producing heat.
At the core of DCIMate is a digital twin that reproduces the thermo-fluid dynamic behavior of the physical environment. It combines discrete CFD simulation for airflow and heat distribution and cooling system models representing CRAC/CRAH units and ventilation systems.
This makes it possible to simulate temperature evolution at the rack level and evaluate the thermal impact of variations in computational workload, in near real time.
Full CFD simulations are too computationally expensive to run continuously at runtime. DCIMate employs a surrogate model trained with machine learning that learns airflow behavior from both physical simulations and real-world data.
The result is near real-time estimation of the thermal field while maintaining good accuracy with respect to the physical model. The surrogate is periodically updated using new operational datasets.
The thermal model is integrated with a workload scheduling pipeline that jointly optimizes IT systems and cooling infrastructure. The system evaluates the thermal impact of different workload distributions via the digital twin and selects the configuration that minimizes total energy consumption, cooling utilization, and thermal hotspots.
Unlike traditional DCIM solutions, DCIMate simultaneously considers server workload, rack temperatures, airflow distribution, and cooling efficiency, enabling joint optimization that was previously not available.
DCIMate operates through a continuous processing pipeline. Sensor data, temperature, airflow, energy logs, and server utilization, is normalized and fed into the digital twin. The surrogate model computes temperature distribution, predicted thermal evolution, and effects of potential workload redistributions within seconds.
The scheduling algorithm then evaluates multiple scenarios and selects the configuration that best balances energy efficiency, thermal stability, and SLA compliance. The system generates workload balancing recommendations, ventilation policies, and optimal cooling configurations. In advanced mode, it operates in closed-loop control, integrating directly with existing DCIM or BMS systems.
DCIMate includes an interactive 3D viewer based on Three.js that represents the data center in near real time, rack layout, room topology, airflow paths and particles, and temperature fields.
The interface facilitates understanding of thermal phenomena and communication of results to operators and non-technical stakeholders. Operators can observe real-time rack temperatures, airflow patterns, and the impact of optimization decisions, all in an intuitive 3D environment.
Each ESP32-based sensor node integrates a low-power microcontroller with Wi-Fi/Ethernet connectivity and hardware interfaces for sensors (I2C, SPI, GPIO). The firmware runs a continuous cycle of sensor data acquisition, local filtering and normalization, and transmission of processed data to the DCIMate platform.
The firmware includes a lightweight embedded agent inspired by zclaw, which exposes hardware functionality through high-level commands. The agent acts as a software interface between DCIMate and the physical sensors, enabling sensor querying, threshold configuration, and firmware updates.
DCIMate is designed for enterprise data centers, HPC infrastructures, edge data centers, and cloud and colocation facilities, wherever energy efficiency drives operational cost and sustainability.