arXiv:2604.11445v1 [cs.DC] 13 Apr 2026
OpenDT: Exploring Datacenter Performance and Sustainability with a Self-Calibrating Digital Twin Radu Nicolae
Jules van der Toorn
Stavriana Kraniti
[email protected] Vrije Universiteit Amsterdam Amsterdam, The Netherlands
[email protected] Vrije Universiteit Amsterdam Amsterdam, The Netherlands
[email protected] Vrije Universiteit Amsterdam Amsterdam, The Netherlands
Houcen Liu
Alexandru Iosup
[email protected] Vrije Universiteit Amsterdam Amsterdam, The Netherlands
[email protected] Vrije Universiteit Amsterdam Amsterdam, The Netherlands
Abstract Datacenters are the backbone of our digital society, but raise numerous operational challenges. We envision digital twins becoming primary instruments in datacenter operations, continuously and autonomously helping with major operational decisions and with adapting ICT infrastructure, live, with a human-in-the-loop. Although fields such as aviation and autonomous driving successfully employ digital twins, an open-source digital twin for datacenters has not been demonstrated to the community. Addressing this challenge, we design, implement, and experiment using OpenDT, an Open-source, Digital Twin for monitoring and operating datacenters through a continuous integration cycle that includes: (1) live and continuous telemetry data; (2) discrete-event simulation using live telemetry from the physical ICT, with self-calibration; and (3) SLO-aware and human-approved feedback to physical ICT. Through trace-driven experiments with a prototype mainly covering stages 1 and 2 of the cycle, we show that (i) OpenDT can be used to reproduce peer-reviewed experiments and extend the analysis with performance and energy-efficiency results; (ii) OpenDT’s online re-calibration can increase digital-twinning accuracy, quantified to a MAPE of 4.39% vs. 7.86% in peer-reviewed work. OpenDT adheres to FAIR/FOSS principles and is available at: https://github.com/atlarge-research/opendt/tree/hcp.
ACM/SPEC International Conference on Performance Engineering (ICPE Companion ’26), May 04–08, 2026, Florence, Italy. ACM, New York, NY, USA, 6 pages. https://doi.org/10.1145/3777911.3800634
B
monitoring and datagen
A Physical ICT Infrastructure
C
E
Digital ICT Infrastructure HITL
SLO-oriented steering
D
Figure 1: High-level overview of datacenter digital twinning: (1) The Physical ICT infrastructure collects telemetry data, (2) the Digital ICT infrastructure ingests telemetry and related data, then (3) mimics (twins) the operation of the Physical ICT, and (4) offers back SLO-aware adjustment feedback. A human-in-the-loop (HITL) oversees the process.
CCS Concepts
1
• Computer systems organization → Cloud computing.
Our digitalized society and economy increasingly rely on digital services running in increasingly larger datacenters [3, 15, 17, 19]. Operators use system analysis, increasingly based on simulation, for designing and operating datacenters [17, 22, 28]. Using simulation as the core of digital twins that mimic datacenter conditions, and support analysis and decision-making in datacenter operations, is an open, emerging challenge [24]. In this work, we design, prototype, and experiment using OpenDT, an Open-source Digital Twin for datacenters. Simulation already supports datacenter operators and scientists with timely and cost-efficient experimentation and analysis [16, 22]. For example, in the European Horizon project Graph Massivizer, simulators predict speedup, failure cost, energy consumption, and CO2-emissions for massive-scale infrastructure [7, 18, 24, 28]. Although simulators are valuable for detailed, realistic analysis of datacenters under workload, in ICT, there is currently no closedloop process that continuously ingests live telemetry, updates the
Keywords OpenDT, datacenters, digital twins, simulation, calibration, performance, sustainability, energy utilization, efficiency. ACM Reference Format: Radu Nicolae, Jules van der Toorn, Stavriana Kraniti, Houcen Liu, and Alexandru Iosup. 2026. OpenDT: Exploring Datacenter Performance and Sustainability with a Self-Calibrating Digital Twin. In Companion of the 17th ∗ First four authors contributed equally to this research. + The fifth author adds conceptual contribution and supervision.
This work is licensed under a Creative Commons Attribution 4.0 International License. ICPE Companion ’26, Florence, Italy © 2026 Copyright held by the owner/author(s). ACM ISBN 979-8-4007-2326-1/2026/05 https://doi.org/10.1145/3777911.3800634
Introduction
ICPE Companion ’26, May 04–08, 2026, Florence, Italy
state inside the simulation, and calibrates the simulator so the multi-metric analysis matches reality. Instead, these steps must be taken independently, with considerable delays and potential errors particularly at boundaries. Figure 1 illustrates the digital twinning process that informed the high-level design of OpenDT. This process involves continuous communication between a physical twin (label A ) and a digital twin ( C ). The physical twin generates telemetry data ( B ), either obtained from the ICT infrastructure’s monitoring and logging systems, or generated from it via advanced analytics or AI-based processes; telemetry data is ingested into the digital twin to be processed directly, or placed in a larger data pool. The digital twin uses techniques to mimic (twin) the operation of the ICT infrastructure, often simulation-based and leveraging SLO-oriented analysis. Closing the loop, the digital twin informs and potentially steers the physical datacenter with SLO-oriented decisions and recommendations. A human-in-the-loop ( E ), aiming to ensure the correct and ethical operation of the process, oversees at least the major decisions and can intervene at any time. In this work, with OpenDT, we make three main contributions: C1. Design and prototype: We design and prototype OpenDT as a datacenter digital twin. The current version of OpenDT supports the cycle of continuous digital-twinning through: (1) telemetry ingestion, (2) trace- and configuration-driven, discrete-event, SLOaware simulation for multi-metric datacenter analysis, (3) support for human-in-the-loop digital twinning. OpenDT is in its early design and prototype stages, with future work in all its main elements, in particular, closing the twinning loop with automated steering (Figure 1, D ). C2. Exploration: We validate OpenDT through real-world experimentation using its prototype. We reproduce and then extend a peer-reviewed experiment [30]. We explore how much OpenDT’s self-calibration improves accuracy. All experiments use real-world workload and energy traces from SURF, the Dutch national supercomputing center [34]. C3. Open Science: We follow the principles of FAIR, FOSS, and reproducible science, and release the engineered OpenDT prototype, together with a reproducibility capsule at: https://github.com/atlarge-research/opendt/tree/hcp .
2
Design of OpenDT: an operational ecosystem for datacenter digital twinning
In this section, we follow the AtLarge design process [20]: first, we identify design requirements based on operational needs, then, present the architecture of OpenDT.
2.1
Requirements analysis
We identify the following design requirements: (FR1) Digital-twinning ICT: OpenDT should ensure active and continuous replication of real-world ICT infrastructure, through a digital twin, continuously updated through telemetry data. (FR2) State-of-the-art, discrete-event simulation: OpenDT should adopt a peer-reviewed, discrete-event simulator, able to predict performance, sustainability, and availability of datacenters under workload, at a user-established granularity.
Radu Nicolae, Jules van der Toorn, Stavriana Kraniti, Houcen Liu, & Alexandru Iosup
Legend
Front-end
Control
Data
B
Decision-making
Simulator G
Interfaces F
Orchestration A
user
Physical ICT P Datacenter Twin
C
SelfCalibrator
H
I
Engine Simulation Engine Engine Engine
Input
Engine Engine Engine Output
Data Management
Virtual ICT V Digital Twin
Orchestrator
Data Platform D Telemetry
E
Figure 2: High-level overview of OpenDT. (FR3) Simulator real-time re-calibration: OpenDT should recalibrate predictions in real-time, based on quantified differences between predictions and real-world measurements. (NFR1) Accurate, ground-truth adjusted predictions: OpenDT should stay within 10% error rate (community-accepted [22, 23, 28, 30]), at ≥90% of the operational time, with dynamic simulation re-calibration (FR3). Inaccurate predictions can influence C-level officers to make wrong decisions [28], and lead to costly downtimes [27], impactful outages [26], and datacenter shutdowns [1]. (NFR2) Performant, lightweight digital twinning: OpenDT must be able to twin 7 days of real-world operation in under 1 hour, measured on a common machine (i.e., not a supercomputer). This supports live decisions for the engineering team. (NFR3) Multi-layer metrics: OpenDT must quantify metrics across performance and sustainability. For each, it must provide at least two metrics. This supports diverse scenarios.
2.2
High-level design of OpenDT
We design for OpenDT the architecture depicted in Figure 2. Overall, OpenDT provides a digital twin state (label V ) that constantly mimics the state of the physical twin ( P ) and is updated with whatif analysis results. The main process supported by this architecture is for closed-loop digital-twinning (FR1), where each iteration (1) begins with telemetry data updates reflecting the physical datacenter’s state ( P ), (2) then, telemetry data is fed into a discrete-event simulator ( H ) for multi-metric datacenter analysis, and into the independent self-calibration process ( G ), and (3) the closed-loop ends with datacenter adjustments, suggested by OpenDT ( G ), and validated and enforced by a human-operator ( A ) for major changes. (Automating the steering process, label D in Figure 1, requires careful interfacing with various kinds of physical ICT resource managers and policies, and is beyond the scope of this work.) A key design choice for OpenDT is the simulator at its core, which is responsible for delivering high-quality predictions that represent as precisely and accurately as needed the ICT infrastructure’s behavior. OpenDT uses the state-of-the-art OpenDC [22], a peer-reviewed, open-source, discrete-event simulator (FR2), with over 7 years of deployment and operation [22, 23, 28]. We augment OpenDC’s predictive model with a real-time calibration process, based on quantified error rate between simulation and reality (FR3). To manage the challenges of the main process, where complexity is compounded by data-intensive telemetry and by computeintensive simulation, OpenDT adopts an orchestrator-centric architecture. The Orchestrator (component C ) operates as a central
OpenDT: Exploring Datacenter Ops with a Self-Calibrating Digital Twin
authority for task scheduling and resource management, and system health monitoring, see Section 2.3. The OpenDT twinning process is based on windows of operation, which are short periods of time during which the system receives telemetry data (asynchronously), runs simulations, processes results, (asynchronously) updates the UI facing the human-in-theloop, and then continues to the next time window. Each window of operation begins with the datacenter twin (Figure 2, label P ), which generates telemetry data across multiple operational layers, either directly observable metrics (e.g., instantaneous power draw, available devices), or derived information (e.g., task progress, operational phenomena), ingested by Telemetry (component D ). OpenDT pre-processes the ingested telemetry, converting it to simulatorready formats and clipping data outside the window of operation, and stores it in the data management component ( E ). Component E links various data-intensive components, and addresses the scale and frequency of telemetry data through a specialized solution—here, a Parquet-based shared file-storage was enough. Continuing the sequence in the window of operation, OpenDT’s orchestrator retrieves the latest collected telemetry data and feeds it into the simulator, where the input (Figure 2, component F ) consists of two main data entries: (1) historical telemetry and past predictions, for simulation calibration ( G ), and (2) the latest telemetry for updating the datacenter state and for predicting future behavior. One or more simulation engines ( H ) begin making predictions, in parallel, and each produces simulation output ( I ). Section 2.4 addresses the interplay between simulation and calibration. The orchestrator ( C ) publishes predictions to the front-end interface ( B ), where the overseeing user ( A ) can intervene in decisionmaking processes. Ultimately, as planned work for a future OpenDT version, the orchestrator ( C ) would redirect the changes to the datacenter twin ( P ), which modifies the physical ICT based on OpenDT’s recommendations. OpenDT supports the human-in-the-loop with several important capabilities. First, users can configure complex scenarios, with specific telemetry data streams and multi-metric analysis; this can be done both off-line, before the twinning process starts, and at runtime, via the UI and API. Second, OpenDT automates complex simulation processes, such as calibration against real-world data as “ground truth” (Section 2.4). Third, OpenDT enables high-complexity techniques that combine individual simulations, e.g., multi-model simulation that combines the results of multiple heterogeneous models, simulated independently [28], to improve accuracy and quantify fine-grained differences in the results produced by each model. We further discuss the design of this component in the technical report [29, §2.7].
2.3
Orchestrator (Component C in Figure 2)
The Orchestrator acts as a central authority and is responsible for managing core operations (e.g., simulation), system monitoring, and ensuring the correctness and coherence of the operational flow of the OpenDT system for each operational cycle. Time-wise, the Orchestrator manages the execution cycle around windows of operation, which are of fixed duration and lead to a lock-step, synchronized schedule. This approach mitigates issues
ICPE Companion ’26, May 04–08, 2026, Florence, Italy
T1
S0
T2 W1
S1
S2
C0
C1
W2 W3
W4
S3 C2
W5
W6
Execution time
Figure 3: Synchronization between simulator (service thread 𝑇 1) and calibrator (𝑇 2). The calibrator runs (events labeled 𝐶), then provides real-time feedback (dashed arrows) to the next simulation run (𝑆), enabling dynamic recalibration based on the incoming workload tasks (𝑊 ). such as data misalignment and ambiguity in simulation step assignments. Without this data, which is produced by the physical twin and does not arrive all at once in the digital twin, situations could become unclear. Similarly, repeatable trace-based simulation requires unambiguous decisions about simulation inputs. During each window, the Orchestrator (Figure 2, component C ) retrieves pre-processed telemetry from the Data Platform (via E ) and starts the simulator with a consistent view of the datacenter state. The Orchestrator also records metadata, such as when a simulation run started and which outputs belong together, which enables correctness and performance analysis of OpenDT itself. OpenDT supports a configurable acceleration factor, expressed as a ratio between simulation and real-world time, with three main modes: (1) Simulation run one-to-one to real time (factor set to 1), (2) Fixed acceleration, e.g., factor set to 10, and (3) Maximum acceleration allowed by computational resources. The first mode is useful for live twinning. The accelerated modes enable faster simulation and exploration of long-term scenarios, but require prior knowledge of the full workload trace and datacenter configuration. In the current implementation, the orchestrator does not manage OpenDT’s own resource allocation, nor the way its components are run. This design choice is deliberate: internal scheduling is delegated to the execution environment, allowing the orchestrator to focus exclusively on validating the digital-twinning loop itself. By keeping its functionality simple, the orchestrator ensures smooth integration among the simulation engine, data platform, and userfacing services while keeping operations interference-free.
2.4
Simulator (H) and Calibrator (G) interplay
OpenDT uses a Simulation Engine ( H ) to inform operational decisions. Even after selecting a state-of-the-art simulator for this purpose, OpenDC [31], simulating large-scale and highly heterogeneous ICT infrastructure with high accuracy, precision, and explainability remains a critical yet non-trivial open scientific challenge in computer systems [11, 22, 28]. Like many state-of-the-art simulators in the field, OpenDC uses a deterministic and static simulation strategy. However, static simulation models can drift and introduce errors as the time horizon increases: hardware behavior varies with temperature, aging, and firmware updates, while workload characteristics evolve over time. These dynamics threaten the validity of assumptions underlying
ICPE Companion ’26, May 04–08, 2026, Florence, Italy
Radu Nicolae, Jules van der Toorn, Stavriana Kraniti, Houcen Liu, & Alexandru Iosup
Table 1: Overview of OpenDT experiments. 𝐶=calibration.
SURF-SARA topology
the static model. To mitigate this problem, OpenDT adds a SelfCalibrator ( G ), which measures the difference between simulationpredicted results and actual telemetry, then continuously adjusts the simulation model to achieve (NFR1). As Figure 3 illustrates, the Simulation Engine (SE) and the SelfCalibrator (SC) run as parallel processes. The SC employs a grid search strategy over the power model’s parameter space: for each calibration cycle, it evaluates a set of candidate parameter values by running short simulations and comparing the results against recent historical telemetry data. The configuration yielding the lowest Mean Absolute Percentage Error (MAPE) is selected and transmitted to the SE for use in subsequent predictions. This pipelined approach (e.g., C0 calibrates S1 in Figure 3) allows calibration to proceed without blocking simulation. In Section 3.4, we demonstrate that this calibration approach improves simulation accuracy both over uncalibrated OpenDT and over state-of-the-art tools [28, 30].
3
Trace-based experiments, enabled by and conducted with OpenDT
We prototype and validate OpenDT. Our experiments (see Table 1) use real-world traces, compare results with a peer-reviewed simulator, and quantify the accuracy gains of real-time recalibration. The experimental results support three main findings (MF1-3): (MF1) OpenDT allows for continuous replay of real-world datacenter operation, with high accuracy. Compared to the measured MAPE error rate of the peer-reviewed simulator FootPrinter 7.86% [30], for this experiment OpenDT achieves a MAPE of 5.13% (§3.3). (MF2) OpenDT’s live self-calibration is enabled by digital twinning, happens regularly and continuously, and improves MAPE; here, from 5.13% to 4.39% (§3.4). (MF3) OpenDT can twin multiple core metrics across operational layers, related to sustainability (e.g., power draw, efficiency, §3.3) and performance (e.g., TFLOPs, §3.3, and CPU utilization [29]).
3.1
Prototype implementation and performance
We implement OpenDT as a Docker Compose microservice system in which the core components (data source, simulator, calibrator, and API) communicate through Apache Kafka [21]. All services mount a shared host directory that serves as a filebased workspace for configuration files and simulation outputs. This shared location makes runtime data easy to inspect and helps with reasoning about the persisted state of the system. For implementation details, see our technical report [29, §3]. Performance: Enabled by the lightweight nature of the OpenDT prototype, we run both experiments on a common off-the-shelf MacBook Pro, with an M1 Max 10-core CPU and 32 GB of RAM. This showcases OpenDT’s capabilities of twinning 7 days of datacenter operation within 46 minutes, thus successfully addressing (NFR2).
C
A
Performance Reports TWIN
OpenDT SIM
Footprinter REAL
B
F
Energy Reports
D SURF-22 workload
TFLOPs (Fig 5b)
E SURF-SARA Datacenter
MAPE (Fig 5a) ?? MAPE (Fig 5a) 3.15% MAPE (Fig 5a) 0.00%
G H
J
(Fig 5c)
E1, §3.3: Reproduce and expand peer-reviewed W, TFLOPs, ✘ experiment with digital twinning capabilities TFLOPs/kWh E2, §3.4: Evaluate self-calibration accuracy W, MAPE [%] ✔
Experimentation Environment
Efficiency Report
𝐶
I
Figure 4: Section 3.3 experiment, adapted and re-run from FootPrinter [30] and extended with OpenDT.
32
Power draw [kW]
Metrics
Efficiency Perfor[TFLOPs mance /kWh] [TFLOPs]
Experiment identifier, section, goals
16
A Ground Truth
0 15000 7500 0 600 300 0
FootPrinter
OpenDT
B C 08/10 09/10 10/10 11/10 12/10 13/10 Time [day/month]
Figure 5: Operational metrics for the compute cluster over time: (A) Power draw simulation results vs. measured reality (reproduced from [28, 30]). (B) Performance measured in TFLOPs. (C) Efficiency measured in TFLOPs/kWh.
3.2
Experiment setup
Infrastructure: In these experiments, we twin the SURF-SARA production cluster at SURF, the Dutch infrastructure for scientific computing. SURF-SARA contains 277 hosts, each with 128 GB of RAM and 16 processing cores running at maximum 2.1 GHz. Workload trace (public): We use SURF-22, a scientific workload trace from SURF, also used in peer-reviewed articles [30, 34]. It traces scientific jobs with an average duration of 39.52 CPU-hours [28]. Quantifying accuracy (error): We employ Mean Absolute Percentage Error (MAPE), a widely used relative-error metric [25, 30, 32], Í 𝑖 calculated as MAPE [%] = 𝑛1 𝑛𝑖=0 𝑅𝑖𝑅−𝑆 × 100, where 𝑛 is the num𝑖 ber of samples, 𝑅 is the real-world data [34], 𝑆 is the simulation data, and 𝑖 is the sample index. Power model: We model CPU power draw adopting the OpenDC [9, 22] analytical formula: 𝑃 (𝑢) = 𝑃idle + (𝑃 max − 𝑃 idle ) (2𝑢 − 𝑢 𝑟 ). Here, 𝑢 is CPU utilization, 𝑃 idle and 𝑃 max represent idle and max power, and 𝑟 is the calibration parameter (see Section 2.4).
3.3
Experiment E1: Peer-reviewed experiment reproduced and extended with OpenDT
Niewenhuis et al. propose FootPrinter [30], a tool for predicting the CO2 footprint of datacenters. Figure 4 illustrates the workflow we
followed in this work, leveraging OpenDT capabilities to reproduce the key experiment [30]. Illustrative for the difference between approaches, whereas FootPrinter ( D ) runs once a hand-tuned energy model [28, 30], OpenDT ( C ) continuously predicts energy consumption at the industry-standard sampling granularity (i.e., 5-minute rate), using a generic predictive model that avoids overfitting for a specific trace. To reproduce the experiment, we use datacenter topology ( A ) and workload trace ( B ) recorded from SURF [28, 30] ( E ), which we regard as “ground-truth” ( I ). We measure the MAPE of OpenDT ( G ), compare it with FootPrinter’s ( H ), and also record datacenter performance data ( I ) that enables us to extend the experiment and also report energy-efficiency results ( J ). The results of this experiment firmly support MF1 and MF3. Accuracy validation by reproducing the peer-reviewed experiment: We measure the accuracy of predictions by determining the MAPE error rate, using the formula described in Section 3.2; the lower the MAPE, the more accurate the prediction. Figure 5A depicts the results of the reproducibility experiment. Between predictions and ground truth (measured reality), we compute FootPrinter’s MAPE as 7.86% and OpenDT’s MAPE as 5.13% (2.73% better). Here, even without simulation recalibration, OpenDT meets the accuracy requirement (NFR1). Extending the peer-reviewed experiment: Beyond quantifying datacenter sustainability, OpenDT can also predict performance and efficiency quickly. Figure 5 depicts the results obtained when we extend the FootPrinter experiment with performance results and demonstrates that OpenDT meets (NFR3). In Figure 5B, we illustrate live, continuous predictions of OpenDT on datacenter performance, which are further processed to produce the efficiency evaluation depicted in Figure 5C. Overall, discretizing OpenDT predictions per hour, we identify the highest efficiency when datacenter performance, quantified in floating-point operations, is the highest. Further investigation on OpenDT’s predictions [29] identifies underutilization of the available infrastructure: during the monitoring period, under 30% of the available processing power is used, while the remaining are idle. Such insights, enabled by OpenDT, could help operators better monitor and plan.
3.4
E2: Evaluating live, self-recalibration against the current simulation state-of-the-art
Addressing the challenge of improving prediction accuracy for digital twins (see Section 2.4), we evaluate the Self-Calibrator. Experimentally, we analyze how real-time recalibration affects OpenDT’s accuracy relative to traditional simulation, which is not live-calibrated. Supporting MF2, Figure 6 depicts the MAPE over time for OpenDT with and without calibration, against the set threshold, i.e., below 10%, 90% of the time (NFR1). Overall, the live-recalibration approach reduces MAPE by 0.74, from 5.13% to 4.39%. We identify for the uncalibrated OpenDT the MAPE < 10% occurs only 86% of the time. However, the live re-calibrated OpenDT achieves MAPE < 10% just over 92% of the time, thus successfully meeting (NFR1). Analysis of calibration vs. no calibration: Although the calibration technique proposed in this work is relatively simple, the calibrated error rate of OpenDT is better than that of the traditional approach. It would be tempting to conclude that using calibration is always beneficial. However, we set out to investigate whether this is the
ICPE Companion ’26, May 04–08, 2026, Florence, Italy
MAPE [%]
OpenDT: Exploring Datacenter Ops with a Self-Calibrating Digital Twin
15
No Calibration
10
NFR Threshold (10%) FootPrinter (7.86%)
With Calibration
5 0
08/10 09/10 10/10 11/10 12/10 13/10 Time [day/month]
Figure 6: Evolution of error rate in power-draw estimation of OpenDT with calibration (MAPE 4.39%), without calibration (MAPE 5.13%), compared with NFR1 and FootPrinter. The horizontal bars at the top distinguish OpenDT over-estimations (solid color) from under (dimmed), without calibration (orange, top bar) and with calibration (green). The black and gray marks at the bottom mark special intervals, see text. case, with results depicted by Figure 6. We observe that there exist significant simulation-periods when the no-calibration technique performs better than the OpenDT calibration (e.g., a window of about 12 hours centered in 11/10, highlighted in black in Figure 6) or equally good (e.g., around 9/10, highlighted in gray). This suggests calibration is important but its technique requires further attention. Prediction behavior: A large body of work, including SPEC RG Cloud’s work on auto-scaling metrics [13], addresses the impact of under- and over-estimation in infrastructure provisioning. Underestimating and under-provisioning ICT infrastructure could lead to performance concerns, system faults, and, ultimately, system failures. In contrast, overprovisioning could lead to sustainability concerns, both environmental and financial, where energy (and thus, money) is wasted by idle infrastructure. In Figure 6, marked by the horizontal bars at the top of the figure, we identify an underestimation trend in OpenDT’s predictive model. This occurs both without calibration (in 85% of discrete-event predictions) and with calibration (66%). Overall, we find calibration alleviates underestimation bias, reducing the frequency of underestimated predictions.
4
Related work: digital twinning and simulation
We now contrast positively OpenDT with prior work. Digital twinning: Digital twins are already widely used in largescale sciences such as aviation and space exploration [4], and enable coarse-grained, dynamic adjustments for the physical twin from a distance, without requiring physical access (e.g., Earth-side mission control of Apollo 13 [2]). Small-scale sciences such as biomedicine use digital twins to enable what-if analysis, e.g., patient monitoring and treatment, where physical access is not viable (e.g., monitoring ventricular activity) using a closed-loop with a doctor-in-theloop [33]. For the medium-scale science of Computer Systems, digital twins are only starting to emerge, mainly due to the inherent intellectual and computational complexity of combining high-level abstractions with detailed system monitoring and massive amounts of telemetry; OpenDT is to-date the first such digital twin to publicly detail its design and become open source.
ICPE Companion ’26, May 04–08, 2026, Florence, Italy
Radu Nicolae, Jules van der Toorn, Stavriana Kraniti, Houcen Liu, & Alexandru Iosup
Simulation: In this work, we leverage the peer-reviewed OpenDC, a discrete-event simulator able to predict datacenter performance, sustainability, and availability, in a time-efficient and cost-effective way [22, 23]. OpenDC can simulate with high explainability, through multiple aligned simulation models [28], and with high robustness, through Meta-Models, which predict by combining predictions of other models, thus alleviating individual model biases [28]. In contrast to other seminal datacenter simulators, such as DCSim [12], CloudSim [5, 10, 14], or SimGrid [6], OpenDT enables a new mode of operation, twinning live and closing the simulation loop, for which it proposes novel design, implementation, and evaluation.
[10] Manoel C. Silva Filho, , et al. 2017. CloudSim Plus: A cloud computing simulation framework pursuing software engineering principles for improved modularity, extensibility and correctness. In IFIP/IEEE. [11] 6G FNS. 2025. Future Network Services: 6G for and by the Netherlands. https: //futurenetworkservices.nl/en/. [12] Sandeep K. S. Gupta, Rose Robin Gilbert, Ayan Banerjee, Zahra Abbasi, Tridib Mukherjee, and Georgios Varsamopoulos. 2011. GDCSim: A tool for analyzing Green Data Center design and resource management techniques. In IGCC. [13] Nikolas Herbst, André Bauer, Samuel Kounev, Giorgos Oikonomou, Erwin Van Eyk, George Kousiouris, Athanasia Evangelinou, Rouven Krebs, Tim Brecht, Cristina L. Abad, and Alexandru Iosup. 2018. Quantifying Cloud Performance and Dependability: Taxonomy, Metric Design and Emerging Challenges. ToMPECS (2018). [14] Tharindu B. Hewage, Shashikant Ilager, Maria Alejandra Rodriguez, and Rajkumar Buyya. 2024. CloudSim express: A novel framework for rapid low code simulation of cloud computing environments. Softw. Pract. Exp. (2024). [15] IDC. 2024. AI Datacenter Capacity, Energy Consumption, and Carbon Emission Projections. https://www.idc.com/getdoc.jsp?containerId=US52131624. [16] Alexandru Iosup. 2024. A VU on Digital Twins to Improve the Performance and Technological Sustainability of Datacenters in the Continuum. In MODSIM. Seattle, USA. [17] Alexandru Iosup, Fernando Kuipers, Ana Lucia Varbanescu, Paola Grosso, Animesh Trivedi, Jan S. Rellermeyer, Lin Wang, Alexandru Uta, and Francesco Regazzoni. 2022. Future Computer Systems and Networking Research in the Netherlands: A Manifesto. CoRR abs/2206.03259 (2022). arXiv:2206.03259 https://doi.org/10.48550/arXiv.2206.03259 [18] Alexandru Iosup, Radu Prodan, Ana Lucia Varbanescu, Sacheendra Talluri, Gilles Magalhaes, Kailhan Hokstam, Hugo Zwaan, Vincent van Beek, Reza Farahani, and Dragi Kimovski. 2023. Graph Greenifier: Towards Sustainable and Energy-Aware Massive Graph Processing in the Computing Continuum. In ICPE. [19] Alexandru Iosup, Alexandru Uta, Laurens Versluis, Georgios Andreadis, Erwin Van Eyk, Tim Hegeman, Sacheendra Talluri, Vincent van Beek, and Lucian Toader. 2018. Massivizing Computer Systems. In ICDCS. [20] Alexandru Iosup, Laurens Versluis, Animesh Trivedi, Erwin Van Eyk, Lucian Toader, Vincent Van Beek, Giulia Frascaria, Ahmed Musaafir, and Sacheendra Talluri. 2019. The AtLarge vision on the design of distributed systems and ecosystems. In ICDCS. IEEE. [21] Jay Kreps. 2011. Kafka : a Distributed Messaging System for Log Processing. https://api.semanticscholar.org/CorpusID:18534081 [22] Fabian Mastenbroek, Georgios Andreadis, Soufiane Jounaid, Wenchen Lai, Jacob Burley, Jaro Bosch, Erwin Van Eyk, Laurens Versluis, Vincent van Beek, and Alexandru Iosup. 2021. OpenDC 2.0: Convenient Modeling and Simulation of Emerging Technologies in Cloud Datacenters. In CCGrid. [23] Fabian Mastenbroek, Tiziano De Matteis, Vincent van Beek, and Alexandru Iosup. 2025. RADiCe: A Risk Analysis Framework for Data Centers. FGCS (2025). [24] Martin Molan, Junaid Ahmed Khan, Andrea Bartolini, Roberta Turra, Giorgio Pedrazzi, Michael Cochez, Alexandru Iosup, Dumitru Roman, Joze M. Rozanec, Ana Lucia Varbanescu, and Radu Prodan. 2023. The Graph-Massivizer Approach Toward a European Sustainable Data Center Digital Twin. In COMPSAC. [25] Juan José Montaño Moreno et al. 2013. Using the R-MAPE index as a resistant measure of forecast accuracy. Psicothema (2013). [26] Sebastian Moss. 2023. Surgeries and procedures paused at Wichita hospitals due to data center outage. Data Center Dynamics, https://www.datacenterdynamics.com/en/news/surgeries-and-procedurespaused-at-wichita-hospitals-due-to-data-center-outage/. [27] NeilMcAllister. 2013. Google goes dark for 2 minutes, kills 40% of world’s net traffic. https://www.theregister.com/2013/08/17/google_outage/. [28] Radu Nicolae, Dante Niewenhuis, Sacheendra Talluri, and Alexandru Iosup. 2026. M3SA: Exploring Datacenter Performance and Climate-Impact with Multi-and Meta-Model Simulation and Analysis. Available at SSRN 5377101 (2026). [29] Radu Nicolae, Jules van der Toorn, Stavriana Kraniti, Houcen Liu, and Alexandru Iosup. 2026. OpenDT: Exploring Datacenter Performance and Sustainability with a Self-Calibrating Digital Twin. Technical Report, https://atlarge-research.com/ pdfs/2026-hcp-opendt-techrep.pdf. [30] Dante Niewenhuis, Sacheendra Talluri, Alexandru Iosup, and Tiziano De Matteis. 2024. FootPrinter: Quantifying Data Center Carbon Footprint. In ICPE. [31] OpenDC Team. 2025. Input: Workload — OpenDC Documentation. https://atlargeresearch.github.io/opendc/docs/documentation/Input/Workload. [32] Oracle. 2024. MAPE. https://docs.oracle.com/en/cloud/saas/planning-budgetingcloud/pfusu/insights_metrics_MAPE.html. [33] Kaan Sel et al. 2024. Building digital twins for cardiovascular health: From principles to clinical impact. Journal of the American Heart Association (2024). [34] Laurens Versluis, Mehmet Çetin, Caspar Greeven, Kristian Laursen, Damian Podareanu, Valeriu Codreanu, Alexandru Uta, and Alexandru Iosup. 2023. Less is not more: We need rich datasets to explore. FGCS (2023).
5
Conclusion and future work
Understanding the performance and sustainability of operating and upcoming datacenters is essential to our society and economy. Addressing the lack of an open-source digital twin for datacenters, in this work we have designed, implemented, and experimented with OpenDT. Overall, our experimental results suggest that OpenDT supports continuous replay of real-world datacenter operations, with high accuracy and, through self-calibration, can improve this accuracy over state-of-the-art approaches. We have released OpenDT as open-source and plan to conduct extensive future experimentation, coupled with real-world infrastructure as closed-loop, as part of a major infrastructure project with over 75 partner institutions [11]. We plan to expand OpenDT to domain-specific operation and couple closed-loop with large-scale ICT infrastructure, running LLM inference workloads. Lastly, we plan to build educational material around digital twinning, aided by OpenDT, and release as open-education, and as optional material in computer systems courses on Computer Organization, Distributed Systems, or in a future edition of the course on Modern Distributed Systems MOOC from edX [8], which already uses a form of OpenDC and could therefore use an exercise based on OpenDT.
Acknowledgment This work is partially supported by EU MSCA CloudStars (101086248) and Horizon Graph Massivizer (101093202), and by the NL National Growth Fund 6G flagship project Future Network Services.
References [1] 2023. Overheating datacenter stopped 2.5 million bank transactions. Laura Dobberstein, The Register, https://www.theregister.com/2023/11/07/overheating_ datacenter_singapore/. [2] B Danette Allen. 2021. Digital twins and living models at NASA. In Digital Twin Summit. [3] Georgios Andreadis, Laurens Versluis, Fabian Mastenbroek, and Alexandru Iosup. 2018. A reference architecture for datacenter scheduling: design, validation, and experiments. In SC. [4] Hakan Aydemir et al. 2020. The Digital Twin Paradigm for Aircraft – Review and Outlook. In AIAA SciTech Forum. AIAA. [5] Rodrigo N. Calheiros et al. 2011. CloudSim: a toolkit for modeling and simulation of cloud computing environments and evaluation of resource provisioning algorithms. SPE (2011). [6] Henri Casanova. 2001. Simgrid: A Toolkit for the Simulation of Application Scheduling. In CCGrid. IEEE Computer Society. [7] Nuria de Lama Sanchez, Peter Haase, Dumitru Roman, and Radu Prodan. 2023. Boosting the Impact of Extreme and Sustainable Graph Processing for Urgent Societal Challenges in Europe Graph-Massivizer. In ICPE. [8] edX Delft University of Technology (DelftX). 2026. Modern Distributed Systems. [9] Xiaobo Fan et al. 2007. Power provisioning for a warehouse-sized computer. In ISCA.