arXiv subjects

Philipp Wiesner

Publications and source records attributed to Philipp Wiesner.

At least 19 recordsLinked to original sources

arXiv2026

Evaluating MFU as a Proxy for GPU Power for Energy-Aware Simulation of LLM Training

High-fidelity performance simulators are essential for designing and configuring efficient AI systems, yet today's tools lack the ability to predict power consumption. Established GPU power models rely on hardware utilization counters, which do not exist until the workload has actually run. This work evaluates whether Model FLOPs Utilization (MFU)-an analytical, software-defined metric relating achieved throughput to peak hardware capability-can serve as a portable, software-defined predictor of GPU power for LLMs. We benchmark almost 3000 single-device training runs across six GPUs, covering different model families, numerical precisions, batch sizes, and context-window lengths. We find that a linear MFU-based power model fits every tested GPU as long as the workload is compute-bound, as in production LLM training. Fitting per-(GPU, dtype, batch size) instead of per-GPU drops the within-cell mean error from around 10% to around 1%, matching the cross-repeat measurement-noise floor.

cs.PF↗

arXiv2026

Exploring Silent Data Corruption as a Reliability Challenge in LLM Training

As Large Language Models (LLMs) scale in size and complexity, the consequences of failures during training become increasingly severe. A major challenge arises from Silent Data Corruption (SDC): hardware-induced faults that bypass system-level detection mechanisms. SDC may behave like benign numerical noise, but can also cause harmful gradient corruption that leads to loss spikes, divergence, or stalled progress. This work provides a controlled study of how intermittent SDC affects LLM pretraining. Using targeted fault injection at the level of GPU matrix-multiply instructions, we characterize the sensitivity of different bit positions, kernel functions, and execution stages. Our analysis shows that locally originating faults can produce impactful corruption, including NaN propagation, short-lived spikes in loss, gradient norm, and attention logits, as well as persistent parameter divergence. Building on the observed corruption signatures, we propose a lightweight detection method that identifies potentially harmful parameter updates. Experiments on LLaMA models with 60M, 350M, and 1.3B parameters demonstrate that recomputing the most recent training step upon detection can effectively mitigate the impact of these events.

cs.LG↗

arXiv2026

Distributed LLM Pretraining During Renewable Curtailment Windows: A Feasibility Study

Training large language models (LLMs) requires substantial compute and energy. At the same time, renewable energy sources regularly produce more electricity than the grid can absorb, leading to curtailment, the deliberate reduction of clean generation that would otherwise go to waste. These periods represent an opportunity: if training is aligned with curtailment windows, LLMs can be pretrained using electricity that is both clean and cheap. This technical report presents a system that performs full-parameter LLM training across geo-distributed GPU clusters during regional curtailment windows, elastically switching between local single-site training and federated multi-site synchronization as sites become available or unavailable. Our prototype trains a 561M-parameter transformer model across three clusters using the Flower federated learning framework, with curtailment periods derived from real-world marginal carbon intensity traces. Preliminary results show that curtailment-aware scheduling preserves training quality while reducing operational emissions to 5-12% of single-site baselines.

cs.DC↗

arXiv2025

Efficiency Will Not Lead to Sustainable Reasoning AI

AI research is increasingly moving toward complex problem solving, where models are optimized not only for pattern recognition but for multi-step reasoning. Historically, computing's global energy footprint has been stabilized by sustained efficiency gains and natural saturation thresholds in demand. But as efficiency improvements are approaching physical limits, emerging reasoning AI lacks comparable saturation points: performance is no longer limited by the amount of available training data but continues to scale with exponential compute investments in both training and inference. This paper argues that efficiency alone will not lead to sustainable reasoning AI and discusses research and policy directions to embed explicit limits into the optimization and governance of such systems.

cs.AI↗

arXiv2025

What happens when nanochat meets DiLoCo?

Although LLM training is typically centralized with high-bandwidth interconnects and large compute budgets, emerging methods target communication-constrained training in distributed environments. The model trade-offs introduced by this shift remain underexplored, and our goal is to study them. We use the open-source nanochat project, a compact 8K-line full-stack ChatGPT-like implementation containing tokenization, pretraining, fine-tuning, and serving, as a controlled baseline. We implement the DiLoCo algorithm as a lightweight wrapper over nanochat's training loop, performing multiple local steps per worker before synchronization with an outer optimizer, effectively reducing communication by orders of magnitude. This inner-outer training is compared against a standard data-parallel (DDP) setup. Because nanochat is small and inspectable, it enables controlled pipeline adaptations and allows direct comparison with the conventional centralized baseline. DiLoCo achieves stable convergence and competitive loss in pretraining but yields worse MMLU, GSM8K, and HumanEval scores after mid-training and SFT. We discover that using DiLoCo-pretrained weights and running mid- and post-training with DDP fails to recover performance, revealing irreversible representation drift from asynchronous updates that impairs downstream alignment. We provide this implementation as an official fork of nanochat on GitHub.

cs.DC↗

arXiv2025

Optimizing Microgrid Composition for Sustainable Data Centers

As computing energy demand continues to grow and electrical grid infrastructure struggles to keep pace, an increasing number of data centers are being planned with colocated microgrids that integrate on-site renewable generation and energy storage. However, while existing research has examined the tradeoffs between operational and embodied carbon emissions in the context of renewable energy certificates, there is a lack of tools to assess how the sizing and composition of microgrid components affects long-term sustainability and power reliability. In this paper, we present a novel optimization framework that extends the computing and energy system co-simulator Vessim with detailed renewable energy generation models from the National Renewable Energy Laboratory's (NREL) System Advisor Model (SAM). Our framework simulates the interaction between computing workloads, on-site renewable production, and energy storage, capturing both operational and embodied emissions. We use a multi-horizon black-box optimization to explore efficient microgrid compositions and enable operators to make more informed decisions when planning energy systems for data centers.

cs.DC↗

arXiv2025

Moving Beyond Marginal Carbon Intensity: A Poor Metric for Both Carbon Accounting and Grid Flexibility

Marginal Carbon Intensity (MCI) has been promoted as an effective metric for carbon-aware computing. Although it is already considered as impractical for carbon accounting purposes, many still view it as valuable when optimizing for grid flexibility by incentivizing electricity usage during curtailment periods. In this statement paper, we argue that MCI is neither reliable nor actionable for either purpose. We outline its fundamental limitations, including non-observability, reliance on opaque predictive models, and the lack of verifiability. Moreover, MCI fails to reflect curtailment caused by high-carbon sources and offers no insight into the quantity of available excess power. We advocate moving beyond MCI and instead call for research on more actionable metrics, such as direct reporting of excess power, explicit modeling of energy storage and grid stability, and integration with emerging granular renewable energy certificate markets.

eess.SY↗

arXiv2025

Quantifying the Energy Consumption and Carbon Emissions of LLM Inference via Simulations

The environmental impact of Large Language Models (LLMs) is rising significantly, with inference now accounting for more than half of their total lifecycle carbon emissions. However, existing simulation frameworks, which are increasingly used to determine efficient LLM deployments, lack any concept of power and, therefore, cannot accurately estimate inference-related emissions. We present a simulation framework to assess the energy and carbon implications of LLM inference under varying deployment setups. First, we extend a high-fidelity LLM inference simulator with a GPU power model that estimates power consumption based on utilization metrics, enabling analysis across configurations like batch size, sequence length, and model parallelism. Second, we integrate simulation outputs into an energy system co-simulation environment to quantify carbon emissions under specific grid conditions and explore the potential of carbon-aware scheduling. Through scenario-based analysis, our framework reveals how inference parameters affect energy demand and carbon footprint, demonstrates a renewable offset potential of up to 69.2% in an illustrative deployment case, and provides a foundation for future carbon-aware inference infrastructure design.

cs.DC↗

arXiv2025

Carbon-Aware Microservice Deployment for Optimal User Experience on a Budget

The carbon footprint of data centers has recently become a critical concern. So far, most carbon-aware strategies have focused on leveraging the flexibility of scheduling decisions for batch processing by shifting the time and location of workload executions. However, such approaches cannot be applied to service-oriented cloud applications, since they have to be reachable at every point in time and often at low latencies. We propose a carbon-aware approach for operating microservices under hourly carbon budgets. By choosing the most appropriate version and horizontal scaleout for each microservice, our strategy maximizes user experience and revenue while staying within budget constraints. Experiments across various application configurations and carbon budgets demonstrate that the approach adapts properly to changing workloads and carbon intensities.

cs.DC↗

arXiv2025

Choosing the Right Battery Model for Data Center Simulations

As demand for computing resources continues to rise, the increasing cost of electricity and anticipated regulations on carbon emissions are prompting changes in data center power systems. Many providers are now operating compute nodes in microgrids, close to renewable power generators and energy storage, to maintain full control over the cost and origin of consumed electricity. Recently, new co-simulation testbeds have emerged that integrate domain-specific simulators to support research, development, and testing of such systems in a controlled environment. Yet, choosing an appropriate battery model for data center simulations remains challenging, as it requires balancing simulation speed, realism, and ease of configuration. In this paper, we implement four different battery models for data center scenarios within the co-simulation framework Vessim and analyze their behavior. The results show that linear models, which consider inefficiencies and power limits, closely match the behavior of complex physics-based models in short-term experiments while offering faster execution, and not requiring knowledge on electrochemical reactions and circuit-level dynamics. In contrast, simple, lossless models fail to accurately represent complex behavior and provide no further runtime advantage.

cs.DC↗

arXiv2025

$\beta$-GNN: A Robust Ensemble Approach Against Graph Structure Perturbation

Graph Neural Networks (GNNs) are playing an increasingly important role in the efficient operation and security of computing systems, with applications in workload scheduling, anomaly detection, and resource management. However, their vulnerability to network perturbations poses a significant challenge. We propose $\beta$-GNN, a model enhancing GNN robustness without sacrificing clean data performance. $\beta$-GNN uses a weighted ensemble, combining any GNN with a multi-layer perceptron. A learned dynamic weight, $\beta$, modulates the GNN's contribution. This $\beta$ not only weights GNN influence but also indicates data perturbation levels, enabling proactive mitigation. Experimental results on diverse datasets show $\beta$-GNN's superior adversarial accuracy and attack severity quantification. Crucially, $\beta$-GNN avoids perturbation assumptions, preserving clean data structure and performance.

cs.LG↗

arXiv2024

Carbon-Aware Quality Adaptation for Energy-Intensive Services

The energy demand of modern cloud services, particularly those related to generative AI, is increasing at an unprecedented pace. To date, carbon-aware computing strategies have primarily focused on batch process scheduling or geo-distributed load balancing. However, such approaches are not applicable to services that require constant availability at specific locations due to latency, privacy, data, or infrastructure constraints. In this paper, we explore how the carbon footprint of energy-intensive services can be reduced by adjusting the fraction of requests served by different service quality tiers. We show that adapting this quality of responses with respect to grid carbon intensity can lead to additional carbon savings beyond resource and energy efficiency. Building on this, we introduce a forecast-based multi-horizon optimization that reaches close-to-optimal carbon savings and is able to automatically adapt service quality for best-effort users to stay within an annual carbon budget. Our approach can reduce the emissions of large-scale LLM services, which we estimate at multiple 10,000 tons of CO2 annually, by up to 10%.

cs.DC↗

arXiv2024

LogRCA: Log-based Root Cause Analysis for Distributed Services

To assist IT service developers and operators in managing their increasingly complex service landscapes, there is a growing effort to leverage artificial intelligence in operations. To speed up troubleshooting, log anomaly detection has received much attention in particular, dealing with the identification of log events that indicate the reasons for a system failure. However, faults often propagate extensively within systems, which can result in a large number of anomalies being detected by existing approaches. In this case, it can remain very challenging for users to quickly identify the actual root cause of a failure. We propose LogRCA, a novel method for identifying a minimal set of log lines that together describe a root cause. LogRCA uses a semi-supervised learning approach to deal with rare and unknown errors and is designed to handle noisy data. We evaluated our approach on a large-scale production log data set of 44.3 million log lines, which contains 80 failures, whose root causes were labeled by experts. LogRCA consistently outperforms baselines based on deep learning and statistical analysis in terms of precision and recall to detect candidate root causes. In addition, we investigated the impact of our deployed data balancing approach, demonstrating that it considerably improves performance on rare failures.

cs.LG↗

arXiv2024

Federated Learning over Connected Modes

Statistical heterogeneity in federated learning poses two major challenges: slow global training due to conflicting gradient signals, and the need of personalization for local distributions. In this work, we tackle both challenges by leveraging recent advances in \emph{linear mode connectivity} -- identifying a linearly connected low-loss region in the parameter space of neural networks, which we call solution simplex. We propose federated learning over connected modes (\textsc{Floco}), where clients are assigned local subregions in this simplex based on their gradient signals, and together learn the shared global solution simplex. This allows personalization of the client models to fit their local distributions within the degrees of freedom in the solution simplex and homogenizes the update signals for the global simplex training. Our experiments show that \textsc{Floco} accelerates the global training process, and significantly improves the local accuracy with minimal computational overhead in cross-silo federated learning settings.

cs.LG↗

arXiv2023

Offloading Real-Time Tasks in IIoT Environments under Consideration of Networking Uncertainties

Offloading is a popular way to overcome the resource and power constraints of networked embedded devices, which are increasingly found in industrial environments. It involves moving resource-intensive computational tasks to a more powerful device on the network, often in close proximity to enable wireless communication. However, many Industrial Internet of Things (IIoT) applications have real-time constraints. Offloading such tasks over a wireless network with latency uncertainties poses new challenges. In this paper, we aim to better understand these challenges by proposing a system architecture and scheduler for real-time task offloading in wireless IIoT environments. Based on a prototype, we then evaluate different system configurations and discuss their trade-offs and implications. Our design showed to prevent deadline misses under high load and network uncertainties and was able to outperform a reference scheduler in terms of successful task throughput. Under heavy task load, where the reference scheduler had a success rate of 5%, our design achieved a success rate of 60%.

cs.NI↗

arXiv2023

Carbon-Awareness in CI/CD

While the environmental impact of digitalization is becoming more and more evident, the climate crisis has become a major issue for society. For instance, data centers alone account for 2.7% of Europe's energy consumption today. A considerable part of this load is accounted for by cloud-based services for automated software development, such as continuous integration and delivery (CI/CD) workflows. In this paper, we discuss opportunities and challenges for greening CI/CD services by better aligning their execution with the availability of low-carbon energy. We propose a system architecture for carbon-aware CI/CD services, which uses historical runtime information and, optionally, user-provided information. We examined the potential effectiveness of different scheduling strategies using real carbon intensity data and 7,392 workflow executions of Github Actions, a popular CI/CD service. Our results show, that user-provided information on workflow deadlines can effectively improve carbon-aware scheduling.

cs.DC↗

arXiv2023

Karasu: A Collaborative Approach to Efficient Cluster Configuration for Big Data Analytics

Selecting the right resources for big data analytics jobs is hard because of the wide variety of configuration options like machine type and cluster size. As poor choices can have a significant impact on resource efficiency, cost, and energy usage, automated approaches are gaining popularity. Most existing methods rely on profiling recurring workloads to find near-optimal solutions over time. Due to the cold-start problem, this often leads to lengthy and costly profiling phases. However, big data analytics jobs across users can share many common properties: they often operate on similar infrastructure, using similar algorithms implemented in similar frameworks. The potential in sharing aggregated profiling runs to collaboratively address the cold start problem is largely unexplored. We present Karasu, an approach to more efficient resource configuration profiling that promotes data sharing among users working with similar infrastructures, frameworks, algorithms, or datasets. Karasu trains lightweight performance models using aggregated runtime information of collaborators and combines them into an ensemble method to exploit inherent knowledge of the configuration search space. Moreover, Karasu allows the optimization of multiple objectives simultaneously. Our evaluation is based on performance data from diverse workload executions in a public cloud environment. We show that Karasu is able to significantly boost existing methods in terms of performance, search time, and cost, even when few comparable profiling runs are available that share only partial common characteristics with the target job.

cs.DC↗

arXiv2023

Towards Benchmarking Power-Performance Characteristics of Federated Learning Clients

Federated Learning (FL) is a decentralized machine learning approach where local models are trained on distributed clients, allowing privacy-preserving collaboration by sharing model updates instead of raw data. However, the added communication overhead and increased training time caused by heterogenous data distributions results in higher energy consumption and carbon emissions for achieving similar model performance than traditional machine learning. At the same time, efficient usage of available energy is an important requirement for battery constrained devices. Because of this, many different approaches on energy-efficient and carbon-efficient FL scheduling and client selection have been published in recent years. However, most of this research oversimplifies power performance characteristics of clients by assuming that they always require the same amount of energy per processed sample throughout training. This overlooks real-world effects arising from operating devices under different power modes or the side effects of running other workloads in parallel. In this work, we take a first look on the impact of such factors and discuss how better power-performance estimates can improve energy-efficient and carbon-efficient FL scheduling.

cs.DC↗