arXiv subjects

Soeren Becker

Publications and source records attributed to Soeren Becker.

At least 19 recordsLinked to original sources

arXiv2026

Beyond Microservices: Testing Web-Scale RCA Methods on GPU-Driven LLM Workloads

Large language model (LLM) services have become an integral part of search, assistance, and decision-making applications. However, unlike traditional web or microservices, the hardware and software stack enabling LLM inference deployment is of higher complexity and far less field-tested, making it more susceptible to failures that are difficult to resolve. Keeping outage costs and quality of service degradations in check depends on shortening mean time to repair, which in practice is gated by how quickly the fault is identified, located, and diagnosed. Automated root cause analysis (RCA) accelerates failure localization by identifying the system component that failed and tracing how the failure propagated. Numerous RCA methods have been developed for traditional services, using request path tracing, resource metric and log data analysis. Yet, existing RCA methods have not been designed for LLM deployments that present distinct runtime characteristics. In this study, we evaluate the effectiveness of RCA methods on a best-practice LLM inference deployment under controlled failure injections. Across 24 methods (20 metric-based, two trace-based, and two multi-source), we find that multi-source approaches achieve the highest accuracy, metric-based methods show fault-type-dependent performance, and trace-based methods largely fail. These results reveal that existing RCA tools do not generalize to LLM systems, motivating tailored analysis techniques and enhanced observability, for which we formulate guidelines.

cs.DC↗

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

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

Distributed Low-Communication Training with Decoupled Momentum Optimization

The training of large models demands substantial computational resources, typically available only in data centers with high-bandwidth interconnects. However, reducing the reliance on high-bandwidth interconnects between nodes enables the use of distributed compute resources as an alternative to centralized data center training. Building on recent advances in distributed model training, we propose an approach that further reduces communication by combining infrequent synchronizations across distributed model replicas with gradient momentum compression. In particular, we treat the optimizer momentum as a signal and decompose the Nesterov momentum into high- and low-frequency components via the discrete cosine transform (DCT). Only the high-frequency components are synchronized across model replicas every $H$ steps. Empirically, our method achieves up to a $16\times$ reduction in communication compared to the baseline DiLoCo, and it generalizes across architectures, including transformer-based language models and convolutional neural networks for images. Overall, this work advances the feasibility of training large models on distributed nodes with low-bandwidth interconnects.

cs.LG↗

arXiv2025

Predicting the Performance of Scientific Workflow Tasks for Cluster Resource Management: An Overview of the State of the Art

Scientific workflow management systems support large-scale data analysis on cluster infrastructures. For this, they interact with resource managers which schedule workflow tasks onto cluster nodes. In addition to workflow task descriptions, resource managers rely on task performance estimates such as main memory consumption and runtime to efficiently manage cluster resources. Such performance estimates should be automated, as user-based task performance estimates are error-prone. In this book chapter, we describe key characteristics of methods for workflow task runtime and memory prediction, provide an overview and a detailed comparison of state-of-the-art methods from the literature, and discuss how workflow task performance prediction is useful for scheduling, energy-efficient and carbon-aware computing, and cost prediction.

cs.DC↗

arXiv2025

WOW: Workflow-Aware Data Movement and Task Scheduling for Dynamic Scientific Workflows

Scientific workflows process extensive data sets over clusters of independent nodes, which requires a complex stack of infrastructure components, especially a resource manager (RM) for task-to-node assignment, a distributed file system (DFS) for data exchange between tasks, and a workflow engine to control task dependencies. To enable a decoupled development and installation of these components, current architectures place intermediate data files during workflow execution independently of the future workload. In data-intensive applications, this separation results in suboptimal schedules, as tasks are often assigned to nodes lacking input data, causing network traffic and bottlenecks. This paper presents WOW, a new scheduling approach for dynamic scientific workflow systems that steers both data movement and task scheduling to reduce network congestion and overall runtime. For this, WOW creates speculative copies of intermediate files to prepare the execution of subsequently scheduled tasks. WOW supports modern workflow systems that gain flexibility through the dynamic construction of execution plans. We prototypically implemented WOW for the popular workflow engine Nextflow using Kubernetes as a resource manager. In experiments with 16 synthetic and real workflows, WOW reduced makespan in all cases, with improvement of up to 94.5% for workflow patterns and up to 53.2% for real workflows, at a moderate increase of temporary storage space. It also has favorable effects on CPU allocation and scales well with increasing cluster size.

cs.DC↗

arXiv2024

Towards a Peer-to-Peer Data Distribution Layer for Efficient and Collaborative Resource Optimization of Distributed Dataflow Applications

Performance modeling can help to improve the resource efficiency of clusters and distributed dataflow applications, yet the available modeling data is often limited. Collaborative approaches to performance modeling, characterized by the sharing of performance data or models, have been shown to improve resource efficiency, but there has been little focus on actual data sharing strategies and implementation in production environments. This missing building block holds back the realization of proposed collaborative solutions. In this paper, we envision, design, and evaluate a peer-to-peer performance data sharing approach for collaborative performance modeling of distributed dataflow applications. Our proposed data distribution layer enables access to performance data in a decentralized manner, thereby facilitating collaborative modeling approaches and allowing for improved prediction capabilities and hence increased resource efficiency. In our evaluation, we assess our approach with regard to deployment, data replication, and data validation, through experiments with a prototype implementation and simulation, demonstrating feasibility and allowing discussion of potential limitations and next steps.

cs.DC↗

arXiv2023

Leveraging Reinforcement Learning for Task Resource Allocation in Scientific Workflows

Scientific workflows are designed as directed acyclic graphs (DAGs) and consist of multiple dependent task definitions. They are executed over a large amount of data, often resulting in thousands of tasks with heterogeneous compute requirements and long runtimes, even on cluster infrastructures. In order to optimize the workflow performance, enough resources, e.g., CPU and memory, need to be provisioned for the respective tasks. Typically, workflow systems rely on user resource estimates which are known to be highly error-prone and can result in over- or underprovisioning. While resource overprovisioning leads to high resource wastage, underprovisioning can result in long runtimes or even failed tasks. In this paper, we propose two different reinforcement learning approaches based on gradient bandits and Q-learning, respectively, in order to minimize resource wastage by selecting suitable CPU and memory allocations. We provide a prototypical implementation in the well-known scientific workflow management system Nextflow, evaluate our approaches with five workflows, and compare them against the default resource configurations and a state-of-the-art feedback loop baseline. The evaluation yields that our reinforcement learning approaches significantly reduce resource wastage compared to the default configuration. Further, our approaches also reduce the allocated CPU hours compared to the state-of-the-art feedback loop by 6.79% and 24.53%.

cs.DC↗

arXiv2023

Towards Advanced Monitoring for Scientific Workflows

Scientific workflows consist of thousands of highly parallelized tasks executed in a distributed environment involving many components. Automatic tracing and investigation of the components' and tasks' performance metrics, traces, and behavior are necessary to support the end user with a level of abstraction since the large amount of data cannot be analyzed manually. The execution and monitoring of scientific workflows involves many components, the cluster infrastructure, its resource manager, the workflow, and the workflow tasks. All components in such an execution environment access different monitoring metrics and provide metrics on different abstraction levels. The combination and analysis of observed metrics from different components and their interdependencies are still widely unregarded. We specify four different monitoring layers that can serve as an architectural blueprint for the monitoring responsibilities and the interactions of components in the scientific workflow execution context. We describe the different monitoring metrics subject to the four layers and how the layers interact. Finally, we examine five state-of-the-art scientific workflow management systems (SWMS) in order to assess which steps are needed to enable our four-layer-based approach.

cs.DC↗

arXiv2023

Macaw: The Machine Learning Magnetometer Calibration Workflow

In Earth Systems Science, many complex data pipelines combine different data sources and apply data filtering and analysis steps. Typically, such data analysis processes are historically grown and implemented with many sequentially executed scripts. Scientific workflow management systems (SWMS) allow scientists to use their existing scripts and provide support for parallelization, reusability, monitoring, or failure handling. However, many scientists still rely on their sequentially called scripts and do not profit from the out-of-the-box advantages a SWMS can provide. In this work, we transform the data analysis processes of a Machine Learning-based approach to calibrate the platform magnetometers of non-dedicated satellites utilizing neural networks into a workflow called Macaw (MAgnetometer CAlibration Workflow). We provide details on the workflow and the steps needed to port these scripts to a scientific workflow. Our experimental evaluation compares the original sequential script executions on the original HPC cluster with our workflow implementation on a commodity cluster. Our results show that through porting, our implementation decreased the allocated CPU hours by 50.2% and the memory hours by 59.5%, leading to significantly less resource wastage. Further, through parallelizing single tasks, we reduced the runtime by 17.5%.

cs.DC↗

arXiv2023

Perona: Robust Infrastructure Fingerprinting for Resource-Efficient Big Data Analytics

Choosing a good resource configuration for big data analytics applications can be challenging, especially in cloud environments. Automated approaches are desirable as poor decisions can reduce performance and raise costs. The majority of existing automated approaches either build performance models from previous workload executions or conduct iterative resource configuration profiling until a near-optimal solution has been found. In doing so, they only obtain an implicit understanding of the underlying infrastructure, which is difficult to transfer to alternative infrastructures and, thus, profiling and modeling insights are not sustained beyond very specific situations. We present Perona, a novel approach to robust infrastructure fingerprinting for usage in the context of big data analytics. Perona employs common sets and configurations of benchmarking tools for target resources, so that resulting benchmark metrics are directly comparable and ranking is enabled. Insignificant benchmark metrics are discarded by learning a low-dimensional representation of the input metric vector, and previous benchmark executions are taken into consideration for context-awareness as well, allowing to detect resource degradation. We evaluate our approach both on data gathered from our own experiments as well as within related works for resource configuration optimization, demonstrating that Perona captures the characteristics from benchmark runs in a compact manner and produces representations that can be used directly.

cs.DC↗

arXiv2023

Probabilistic Time Series Forecasting for Adaptive Monitoring in Edge Computing Environments

With increasingly more computation being shifted to the edge of the network, monitoring of critical infrastructures, such as intermediate processing nodes in autonomous driving, is further complicated due to the typically resource-constrained environments. In order to reduce the resource overhead on the network link imposed by monitoring, various methods have been discussed that either follow a filtering approach for data-emitting devices or conduct dynamic sampling based on employed prediction models. Still, existing methods are mainly requiring adaptive monitoring on edge devices, which demands device reconfigurations, utilizes additional resources, and limits the sophistication of employed models. In this paper, we propose a sampling-based and cloud-located approach that internally utilizes probabilistic forecasts and hence provides means of quantifying model uncertainties, which can be used for contextualized adaptations of sampling frequencies and consequently relieves constrained network resources. We evaluate our prototype implementation for the monitoring pipeline on a publicly available streaming dataset and demonstrate its positive impact on resource efficiency in a method comparison.

cs.DC↗

arXiv2022

Federated Learning for Autoencoder-based Condition Monitoring in the Industrial Internet of Things

Enabled by the increasing availability of sensor data monitored from production machinery, condition monitoring and predictive maintenance methods are key pillars for an efficient and robust manufacturing production cycle in the Industrial Internet of Things. The employment of machine learning models to detect and predict deteriorating behavior by analyzing a variety of data collected across several industrial environments shows promising results in recent works, yet also often requires transferring the sensor data to centralized servers located in the cloud. Moreover, although collaborating and sharing knowledge between industry sites yields large benefits, especially in the area of condition monitoring, it is often prohibited due to data privacy issues. To tackle this situation, we propose an Autoencoder-based Federated Learning method utilizing vibration sensor data from rotating machines, that allows for a distributed training on edge devices, located on-premise and close to the monitored machines. Preserving data privacy and at the same time exonerating possibly unreliable network connections of remote sites, our approach enables knowledge transfer across organizational boundaries, without sharing the monitored data. We conducted an evaluation utilizing two real-world datasets as well as multiple testbeds and the results indicate that our method enables a competitive performance compared to previous results, while significantly reducing the resource and network utilization.

cs.LG↗

arXiv2022

Network Emulation in Large-Scale Virtual Edge Testbeds: A Note of Caution and the Way Forward

The growing research and industry interest in the Internet of Things and the edge computing paradigm has increased the need for cost-efficient virtual testbeds for large-scale distributed applications. Researchers, students, and practitioners need to test and evaluate the interplay of hundreds or thousands of real software components and services connected with a realistic edge network without access to physical infrastructure. While advances in virtualization technologies have enabled parts of this, network emulation as a crucial part in the development of edge testbeds is lagging behind: As we show in this paper, NetEm, the current state-of-the-art network emulation tooling included in the Linux kernel, imposes prohibitive scalability limits. We quantify these limits, investigate possible causes, and present a way forward for network emulation in large-scale virtual edge testbeds based on eBPFs.

cs.DC↗

arXiv2022

SyncMesh: Improving Data Locality for Function-as-a-Service in Meshed Edge Networks

The increasing use of Internet of Things devices coincides with more communication and data movement in networks, which can exceed existing network capabilities. These devices often process sensor or user information, where data privacy and latency are a major concern. Therefore, traditional approaches like cloud computing do not fit well, yet new architectures such as edge computing address this gap. In addition, the Function-as-a-Service (FaaS) paradigm gains in prevalence as a workload execution platform, however the decoupling of storage results in further challenges for highly distributed edge environments. To address this, we propose SyncMesh, a system to manage, query, and transform data in a scalable and stateless manner by leveraging the capabilities of Function-as-a-Service and at the same time enabling data locality. Furthermore, we provide a prototypical implementation and evaluate it against established centralized and decentralized systems in regard to traffic usage and request times. The preliminary results indicate that SyncMesh is able to exonerate the network layer and accelerate the transmission of data to clients, while simultaneously improving local data processing.

cs.DC↗

arXiv2022

Efficient Runtime Profiling for Black-box Machine Learning Services on Sensor Streams

In highly distributed environments such as cloud, edge and fog computing, the application of machine learning for automating and optimizing processes is on the rise. Machine learning jobs are frequently applied in streaming conditions, where models are used to analyze data streams originating from e.g. video streams or sensory data. Often the results for particular data samples need to be provided in time before the arrival of next data. Thus, enough resources must be provided to ensure the just-in-time processing for the specific data stream. This paper focuses on proposing a runtime modeling strategy for containerized machine learning jobs, which enables the optimization and adaptive adjustment of resources per job and component. Our black-box approach assembles multiple techniques into an efficient runtime profiling method, while making no assumptions about underlying hardware, data streams, or applied machine learning jobs. The results show that our method is able to capture the general runtime behaviour of different machine learning jobs already after a short profiling phase.

cs.DC↗

arXiv2021

LOS: Local-Optimistic Scheduling of Periodic Model Training For Anomaly Detection on Sensor Data Streams in Meshed Edge Networks

Anomaly detection is increasingly important to handle the amount of sensor data in Edge and Fog environments, Smart Cities, as well as in Industry 4.0. To ensure good results, the utilized ML models need to be updated periodically to adapt to seasonal changes and concept drifts in the sensor data. Although the increasing resource availability at the edge can allow for in-situ execution of model training directly on the devices, it is still often offloaded to fog devices or the cloud. In this paper, we propose Local-Optimistic Scheduling (LOS), a method for executing periodic ML model training jobs in close proximity to the data sources, without overloading lightweight edge devices. Training jobs are offloaded to nearby neighbor nodes as necessary and the resource consumption is optimized to meet the training period while still ensuring enough resources for further training executions. This scheduling is accomplished in a decentralized, collaborative and opportunistic manner, without full knowledge of the infrastructure and workload. We evaluated our method in an edge computing testbed on real-world datasets. The experimental results show that LOS places the training executions close to the input sensor streams, decreases the deviation between training time and training period by up to 40% and increases the amount of successfully scheduled training jobs compared to an in-situ execution.

cs.DC↗

arXiv2021

EdgePier: P2P-based Container Image Distribution in Edge Computing Environments

Edge and fog computing architectures utilize container technologies in order to offer a lightweight application deployment. Container images are stored in registry services and operated by orchestration platforms to download and start the respective applications on nodes of the infrastructure. During large application rollouts, the connection to the registry is prone to become a bottleneck, which results in longer provisioning times and deployment latencies. Previous work has mainly addressed this problem by proposing scalable registries, leveraging the BitTorrent protocol or distributed storage to host container images. However, for lightweight and dynamic edge environments the overhead of several dedicated components is not feasible in regard to its interference of the actual workload and is subject to failures due to the introduced complexity. In this paper we introduce a fully decentralized container registry called EdgePier, that can be deployed across edge sites and is able to decrease container deployment times by utilizing peer-to-peer connections between participating nodes. Image layers are shared without the need for further centralized orchestration entities. The conducted evaluation shows that the provisioning times are improved by up to 65% in comparison to a baseline registry, even with limited bandwidth to the cloud.

cs.DC↗