Vessim: Carbon-Aware Co-Simulation Testbed

• Vessim is a co-simulation testbed designed for carbon-aware computing, modeling interactions between computing workloads and energy subsystems like microgrids.
• It integrates domain-specific simulators, historical datasets, and real hardware through the Mosaik framework for reproducible testing and rapid prototyping.
• The testbed supports both accelerated offline evaluation and real-time hardware-in-the-loop experiments, standardizing research on energy-aware applications.

Searching arXiv for Vessim and closely related extensions to ground the article in the cited literature. Vessim is a co-simulation testbed for carbon-aware applications and systems that models the interaction between computing workloads and energy systems, particularly microgrids with renewable generation and storage, without requiring a full physical test environment (Wiesner et al., 2023). It is positioned as a response to a methodological bottleneck in carbon-aware computing: most prior evaluations rely on custom, one-off simulation environments, which hampers comparability, reproducibility, and realistic validation (Wiesner et al., 2023). Built on the Mosaik co-simulation framework, Vessim provides a modular and extensible environment that can connect domain-specific simulators, historical datasets, and real software or hardware components, thereby supporting research, rapid prototyping, reproducible testing, continuous testing, and digital-twin-oriented operation (Wiesner et al., 2023).

1. Origin and motivating problem

The development of Vessim is motivated by two concurrent trends. First, computing demand is rising, especially due to machine learning, big data, and large-scale datacenters (Wiesner et al., 2023). Second, carbon-aware computing requires better interfaces to energy systems, because signals such as solar generation, battery state of charge, and grid carbon intensity are often hidden behind abstraction layers (Wiesner et al., 2023). The underlying premise is that applications and system software cannot align demand with clean-energy availability unless those energy-system states are observable and actionable.

The problem formulation in the original paper is methodological as much as infrastructural. Research on energy-aware and carbon-aware applications, and on the interfaces between computing and energy systems, is described as complex because suitable testing environments are scarce (Wiesner et al., 2023). Existing work typically depends on bespoke simulation setups, leading to repeated implementation effort and limited comparability of results (Wiesner et al., 2023). Vessim is therefore proposed not merely as another simulator, but as a reusable co-simulation environment intended to standardize experimentation at the boundary between computing and energy systems.

The scope of that ambition extends beyond academic evaluation. The paper explicitly states that Vessim is meant to accelerate research on carbon-aware systems, rapid prototyping of new application ideas, reproducible testing, continuous integration / continuous testing, and digital twin use cases for operational decision support (Wiesner et al., 2023). This suggests a conception of carbon-aware computing as an end-to-end systems problem in which development, testing, and operation all require shared experimental substrates.

2. Co-simulation model and architectural basis

Vessim is implemented on top of Mosaik, which serves as the underlying co-simulation framework (Wiesner et al., 2023). In this arrangement, Vessim acts as a testbed layer on top of Mosaik, defining common interfaces and utilities tailored to carbon-aware computing (Wiesner et al., 2023). Mosaik contributes APIs for connecting simulators written in different languages, support for both continuous and discrete-event simulators, TCP-based integration of real systems, scenario descriptions for subsystem interaction, and runtime coordination of synchronization and message passing (Wiesner et al., 2023).

The choice of co-simulation is central. In Vessim’s formulation, a larger system is decomposed into subsystems that can each be simulated by specialized tools or represented by real components (Wiesner et al., 2023). This matches the structure of energy-aware computing deployments, where renewable generation, storage, power-flow behavior, and compute demand are all distinct but interdependent subsystems. Rather than forcing them into a single monolithic model, Vessim composes them through coordinated simulation.

Later work further clarifies how this architecture is used in microgrid-oriented data center studies. A Vessim microgrid is described as comprising four stepped subsystems: Actors, Grid, Controllers, and Storage (Kilian et al., 21 Jun 2025). Actors model generators and consumers such as solar panels and server loads; the Grid performs power-flow analysis and computes power imbalance; Controllers represent user-defined policies or applications; and Storage models the battery system (Kilian et al., 21 Jun 2025). In that formulation, the simulation step at time $t$ proceeds by computing the power delta $\Delta p_t$, applying a policy decision for storage power and duration, updating the battery state, and then calculating net grid exchange as

$e_t = \Delta p_t \cdot d_t - e_t^s$

where $e_t^s$ is the actual stored or discharged energy returned by the battery model (Kilian et al., 21 Jun 2025).

This stepped architecture shows that Vessim is not restricted to trace replay. It supports closed-loop interaction in which controllers and applications can observe system state and affect subsequent energy flows. A plausible implication is that Vessim is intended to evaluate control policies and system interfaces under realistic coupling, not only to measure passive demand against exogenous energy traces.

3. Simulated components and data interfaces

Vessim is designed to connect domain-specific simulators for renewable generation, storage, and power systems, along with external data sources (Wiesner et al., 2023). For renewable power generation, the original paper identifies solar power through replay of historical solar production data or weather-derived irradiance data, wind power via windpowerlib or the System Advisor Model (SAM), and additional renewable or low-carbon sources including concentrating solar power, marine energy, biomass, geothermal, and potentially static low-carbon sources such as nuclear (Wiesner et al., 2023).

For storage, the paper lists battery models implemented through PyBaMM, simplified models such as the C/L/C model, battery degradation models from SAM, hydrogen storage through fuel-cell and electrolyzer-related models, and other storage types such as flywheels and supercapacitors (Wiesner et al., 2023). For power-system and microgrid analysis, Vessim can integrate tools such as MATPOWER, pandapower, and PyPSA when detailed power-flow or optimal power-flow modeling is required (Wiesner et al., 2023).

The framework also supports external data and forecast integration. It can emulate or connect to weather services, solar and power-production forecasts, carbon intensity APIs, historical data traces, and potentially live providers such as Electricity Maps, WattTime, and Solcast (Wiesner et al., 2023). This data-orientation is important because carbon-aware evaluation depends on temporal variation in both energy supply and grid emissions. Vessim correspondingly supports replaying historical power production traces, replaying carbon intensity data, replaying or emulating forecasts, and replaying workload traces from Alibaba, Azure, and Google (Wiesner et al., 2023).

A later extension substantially increases the realism of renewable modeling by integrating NREL’s System Advisor Model into Vessim (Irion et al., 6 Aug 2025). In that work, the new component instantiates and runs a SAM simulation, extracts the resulting power time series, and serves those time-indexed power values to Vessim actors during simulation (Irion et al., 6 Aug 2025). The extension adds high-resolution solar generation modeling, wind generation modeling, battery storage modeling, and direct coupling of SAM outputs into Vessim’s actor-signal architecture (Irion et al., 6 Aug 2025). This allows Vessim to simulate a microgrid in which workload demand, renewable supply, and battery behavior interact dynamically over time with more realistic generation models than simple synthetic profiles (Irion et al., 6 Aug 2025).

4. Software-in-the-loop, hardware-in-the-loop, and temporal execution modes

A defining feature of Vessim is that it can integrate real software and real hardware, not only simulated energy-system components (Wiesner et al., 2023). In software-in-the-loop experiments, real applications run inside the testbed while Vessim simulates the surrounding energy system (Wiesner et al., 2023). The framework periodically queries the computing infrastructure’s power demand, computes energy flows, and determines how demand is served by on-site generation, batteries, or the public grid (Wiesner et al., 2023). This makes it possible to test application behavior under controlled energy conditions while preserving realistic control and observability.

Hardware-in-the-loop support extends this principle to physical nodes. The paper mentions a Raspberry Pi 3b+ with current and voltage measured by an INA219 DC sensor, as well as a virtual node on Google Cloud with a software power model (Wiesner et al., 2023). For real hardware, external measurement is preferred when possible; for virtualized infrastructure, power demand can be estimated from utilization metrics and power models (Wiesner et al., 2023). This hybrid capability is important because carbon-aware control often depends on deployment-specific interactions that are difficult to approximate purely in software.

The paper makes a strong distinction between simulation-time and real-time operation. If everything is simulated, Vessim can run much faster than real time, enabling months or years of simulated time to be explored quickly (Wiesner et al., 2023). If real software or hardware participates, however, the simulation must operate in real time, so each time step must complete before wall-clock time advances beyond it (Wiesner et al., 2023). To stabilize real-time co-simulation, Vessim’s API server includes a cache that aggregates incoming SET requests, forwards them to the simulation at defined intervals, and serves GET requests directly from periodically updated cached state (Wiesner et al., 2023). This reduces the risk that the simulation falls behind.

The distinction is methodologically significant. Purely simulated studies support long-horizon evaluation over historical traces, whereas hybrid experiments prioritize fidelity of interaction with real artifacts. This suggests that Vessim is designed to span two modes of systems research: accelerated offline evaluation and synchronized online experimentation.

5. Interfaces, extensibility, and control surface

Extensibility is a core design principle of Vessim. The original paper states that users can configure a partially or fully simulated energy system and integrate different kinds of components depending on the experiment, including simulated subsystems, real software, and real hardware (Wiesner et al., 2023). This is not only a software-engineering property but also a research requirement, because carbon-aware systems may need to evaluate different combinations of energy, workload, and control subsystems.

Vessim provides common interfaces intended to simplify experiment configuration and system interaction (Wiesner et al., 2023). A notable mechanism is the REST API, which exposes visibility into energy state and control capabilities (Wiesner et al., 2023). Examples of exposed or configurable interactions include current carbon intensity, solar production, battery charge/discharge rates, and throttling application power usage (Wiesner et al., 2023). The API and the exposed information are described as fully configurable by users (Wiesner et al., 2023).

Later work on battery modeling further formalizes interface boundaries. The 2025 battery-model paper adds a Storage object that standardizes the battery interface and a MicrogridPolicy object that decides how much of the grid imbalance should be absorbed by the battery versus exchanged with the public grid (Kilian et al., 21 Jun 2025). The storage layer exposes state of charge and potentially other model-specific state variables, while controllers can update storage or policy parameters during runtime (Kilian et al., 21 Jun 2025). The paper emphasizes both visibility and control: SoC is the main indicator of battery energy availability, and custom internal state may also be passed to controllers through Mosaik (Kilian et al., 21 Jun 2025).

These extensions illustrate how Vessim’s interface model separates plant dynamics, policy logic, and application behavior. This suggests that the framework is well-suited to studies in which the central question concerns what information should be exposed to software and how that software should act on it under different abstraction levels.

6. Representative extensions and experimental uses

The original Vessim paper is explicitly characterized as a vision and prototype paper rather than a full benchmark study (Wiesner et al., 2023). It reports a working prototype that already includes solar power production replay, grid carbon intensity forecast replay, a simple battery model, and interaction with real applications on virtual and physical hosts (Wiesner et al., 2023). The evaluation is therefore framed as a demonstration of feasibility.

Subsequent work extends Vessim in two directions: planning-oriented microgrid optimization and high-fidelity battery-model comparison. The microgrid-composition paper presents an optimization framework that combines Vessim co-simulation, SAM-based renewable models, black-box multi-objective optimization with Optuna, and carbon accounting for both operational and embodied emissions (Irion et al., 6 Aug 2025). The framework optimizes over the number of wind turbines, installed solar panel area or solar capacity, and battery storage capacity, and returns a Pareto front of non-dominated solutions (Irion et al., 6 Aug 2025). In the reported case study, the full parameter space contains 1089 valid combinations; exhaustive search exceeds 24 hours, while black-box search uses 350 trials with population size 50, recovers about 80% of Pareto-optimal solutions, and yields roughly a 2.4× speed-up (Irion et al., 6 Aug 2025).

That work also demonstrates Vessim’s use with site-specific workload and resource traces. It uses Perlmutter workload traces for Berkeley, CA and Houston, TX, with average power around 1.62 MW during the experimental window; solar data come from NSRDB and wind data from the NREL WIND Toolkit (Irion et al., 6 Aug 2025). Operational emissions are measured in $\text{tCO}_2/\text{day}$ under a GHG Protocol Scope 2 perspective using average carbon intensity data from Electricity Maps, while embodied emissions are treated as a one-time upfront investment in $\text{tCO}_2$ under the Scope 3 treatment of capital goods (Irion et al., 6 Aug 2025). The paper uses Vessim to estimate renewable coverage, battery cycling, and the coupled trade-off between infrastructure investment and long-term emissions (Irion et al., 6 Aug 2025).

The battery-model paper extends Vessim to support step-wise execution of external battery simulators so that different storage models can be compared under identical scenarios (Kilian et al., 21 Jun 2025). It implements four battery models within Vessim—SimpleBattery, CLCBattery, PybammBattery, and LiionBatteryPack—and evaluates them in constant-power charging, constant-power discharging, and a simplified two-day microgrid scenario (Kilian et al., 21 Jun 2025). The scenario includes two server nodes consuming 200 W and 50 W, total consumption multiplied by PUE $= 1.3$, a $10m^2$ solar panel with 15% efficiency, solar radiation data from Berlin in June 2021, a minimum SoC of 30%, and a controller that forces a battery charge at 3200 W from 11 AM to 12 AM on the second day (Kilian et al., 21 Jun 2025). A principal result is that the linear CLCBattery model closely matches the behavior of more complex physics-based models in short-term experiments while adding almost no runtime overhead relative to the simplest model (Kilian et al., 21 Jun 2025).

These extensions show that Vessim has evolved into a common co-simulation substrate for studies that differ substantially in objective: controller evaluation, microgrid planning, and component-model fidelity. A plausible implication is that the framework’s main contribution lies less in any single simulator implementation than in the stable composition layer that permits such heterogeneous studies.

7. Research significance, use cases, and limitations

The use cases identified in the original paper span both scientific and operational settings. They include energy system composition, such as evaluating what happens when solar panels and batteries are added to an edge node; energy system abstractions, including safe, software-defined, or virtualized energy interfaces; energy system interfaces, including new component integration, handling weather and forecast sources, geo-distributed coordination, and access control and security; carbon-aware applications using realistic simulations and common datasets; and development and operation through continuous testing, rollout validation, digital twins, and risk assessment under extreme events such as power outages (Wiesner et al., 2023).

The framework’s broader significance lies in its attempt to move carbon-aware computing away from ad hoc, isolated simulations and toward shared, reproducible, and realistic experimentation (Wiesner et al., 2023). By exposing configurable energy-system interfaces and supporting both historical replay and hybrid experiments with real systems, Vessim creates conditions under which workload shifting, energy-aware control, microgrid sizing, and storage-policy behavior can be studied within a common methodological scaffold.

At the same time, the papers emphasize that Vessim is not a finished production platform. The original paper states that the current prototype is still evolving and identifies future work in validating some preconfigured environments against real energy systems to understand limitations and threats to validity (Wiesner et al., 2023). This qualification matters because co-simulation realism depends on the fidelity of the connected component models, the accuracy of forecast emulation, and the correspondence between simulated interfaces and real operational constraints.

A recurring misconception would be to treat Vessim as a single fixed simulator for a narrow problem class. The literature instead presents it as a co-simulation environment that can host a range of models and experimental designs, from software-in-the-loop application testing to full-year microgrid composition studies and comparative battery-model analysis (Wiesner et al., 2023, Irion et al., 6 Aug 2025, Kilian et al., 21 Jun 2025). Another misconception would be to assume that Vessim inherently guarantees realism; the papers instead indicate that realism is contingent on the selected simulators, data sources, and validation against real systems (Wiesner et al., 2023).

In that sense, Vessim’s main contribution is infrastructural and methodological. It provides a functioning prototype and an extensible framework that bridges computing and energy-system simulation, supports real software and hardware integration, exposes configurable interfaces, and encourages common datasets and reusable components as a foundation for more reproducible carbon-aware computing research (Wiesner et al., 2023).