arXiv subjects

Stefan Schmid

Publications and source records attributed to Stefan Schmid.

At least 73 records · Page 4Linked to original sources

arXiv2025

Predicting the Road Ahead: A Knowledge Graph based Foundation Model for Scene Understanding in Autonomous Driving

The autonomous driving field has seen remarkable advancements in various topics, such as object recognition, trajectory prediction, and motion planning. However, current approaches face limitations in effectively comprehending the complex evolutions of driving scenes over time. This paper proposes FM4SU, a novel methodology for training a symbolic foundation model (FM) for scene understanding in autonomous driving. It leverages knowledge graphs (KGs) to capture sensory observation along with domain knowledge such as road topology, traffic rules, or complex interactions between traffic participants. A bird's eye view (BEV) symbolic representation is extracted from the KG for each driving scene, including the spatio-temporal information among the objects across the scenes. The BEV representation is serialized into a sequence of tokens and given to pre-trained language models (PLMs) for learning an inherent understanding of the co-occurrence among driving scene elements and generating predictions on the next scenes. We conducted a number of experiments using the nuScenes dataset and KG in various scenarios. The results demonstrate that fine-tuned models achieve significantly higher accuracy in all tasks. The fine-tuned T5 model achieved a next scene prediction accuracy of 86.7%. This paper concludes that FM4SU offers a promising foundation for developing more comprehensive models for scene understanding in autonomous driving.

cs.CL↗

arXiv2025

Dynamic Demand-Aware Link Scheduling for Reconfigurable Datacenters

Emerging reconfigurable datacenters allow to dynamically adjust the network topology in a demand-aware manner. These datacenters rely on optical switches which can be reconfigured to provide direct connectivity between racks, in the form of edge-disjoint matchings. While state-of-the-art optical switches in principle support microsecond reconfigurations, the demand-aware topology optimization constitutes a bottleneck. This paper proposes a dynamic algorithms approach to improve the performance of reconfigurable datacenter networks, by supporting faster reactions to changes in the traffic demand. This approach leverages the temporal locality of traffic patterns in order to update the interconnecting matchings incrementally, rather than recomputing them from scratch. In particular, we present six (batch-)dynamic algorithms and compare them to static ones. We conduct an extensive empirical evaluation on 176 synthetic and 39 real-world traces, and find that dynamic algorithms can both significantly improve the running time and reduce the number of changes to the configuration, especially in networks with high temporal locality, while retaining matching weight.

cs.NI↗

arXiv2025

Ethereal: Divide and Conquer Network Load Balancing in Large-Scale Distributed Training

Large-scale distributed training in production datacenters constitutes a challenging workload bottlenecked by network communication. In response, both major industry players (e.g., Ultra Ethernet Consortium) and parts of academia have surprisingly, and almost unanimously, agreed that packet spraying is \emph{necessary} to improve the performance of large-scale distributed training workloads. In this paper, we challenge this prevailing belief and pose the question: \emph{How close can singlepath transport come to matching the performance of packet spraying?} We demonstrate that singlepath transport (from a NIC's perspective) is sufficient and can perform nearly as well as ideal packet spraying, particularly in the context of distributed training in CLOS-based topologies. Our assertion is based on four key observations about workloads driven by collective communication patterns: \emph{(i)} flow sizes are known upon arrival, \emph{(ii)} flow sizes are equal within each step of a collective, \emph{(iii)} the completion time of a collective is more critical than individual flow completion times, and \emph{(iv)} flows can be \emph{split} upon arrival to control load balancing directly from the application layer. We present Ethereal, a simple distributed load balancing algorithm that opportunistically splits flows and assigns paths to each flow in a transparent manner, requiring little to no changes to existing RDMA NICs. Our evaluation, spanning a wide range of collective communication algorithms and GPT models using Astra-Sim, shows that Ethereal significantly reduces the completion times by up to $30\%$ compared to packet spraying and by up to $40\%$ compared to REPS, even under link failures. This paper offers an alternative perspective for developing next-generation transport protocols tailored to large-scale distributed training.

cs.NI↗

arXiv2025

Revolutionizing Datacenter Networks via Reconfigurable Topologies

With the popularity of cloud computing and data-intensive applications such as machine learning, datacenter networks have become a critical infrastructure for our digital society. Given the explosive growth of datacenter traffic and the slowdown of Moore's law, significant efforts have been made to improve datacenter network performance over the last decade. A particularly innovative solution is reconfigurable datacenter networks (RDCNs): datacenter networks whose topologies dynamically change over time, in either a demand-oblivious or a demand-aware manner. Such dynamic topologies are enabled by recent optical switching technologies and stand in stark contrast to state-of-the-art datacenter network topologies, which are fixed and oblivious to the actual traffic demand. In particular, reconfigurable demand-aware and 'self-adjusting' datacenter networks are motivated empirically by the significant spatial and temporal structures observed in datacenter communication traffic. This paper presents an overview of reconfigurable datacenter networks. In particular, we discuss the motivation for such reconfigurable architectures, review the technological enablers, and present a taxonomy that classifies the design space into two dimensions: static vs. dynamic and demand-oblivious vs. demand-aware. We further present a formal model and discuss related research challenges. Our article comes with complementary video interviews in which three leading experts, Manya Ghobadi, Amin Vahdat, and George Papen, share with us their perspectives on reconfigurable datacenter networks.

cs.NI↗

arXiv2025

Robust Visual Representation Learning with Multi-modal Prior Knowledge for Image Classification Under Distribution Shift

Despite the remarkable success of deep neural networks (DNNs) in computer vision, they fail to remain high-performing when facing distribution shifts between training and testing data. In this paper, we propose Knowledge-Guided Visual representation learning (KGV) - a distribution-based learning approach leveraging multi-modal prior knowledge - to improve generalization under distribution shift. It integrates knowledge from two distinct modalities: 1) a knowledge graph (KG) with hierarchical and association relationships; and 2) generated synthetic images of visual elements semantically represented in the KG. The respective embeddings are generated from the given modalities in a common latent space, i.e., visual embeddings from original and synthetic images as well as knowledge graph embeddings (KGEs). These embeddings are aligned via a novel variant of translation-based KGE methods, where the node and relation embeddings of the KG are modeled as Gaussian distributions and translations, respectively. We claim that incorporating multi-model prior knowledge enables more regularized learning of image representations. Thus, the models are able to better generalize across different data distributions. We evaluate KGV on different image classification tasks with major or minor distribution shifts, namely road sign classification across datasets from Germany, China, and Russia, image classification with the mini-ImageNet dataset and its variants, as well as the DVM-CAR dataset. The results demonstrate that KGV consistently exhibits higher accuracy and data efficiency across all experiments.

cs.CV↗

arXiv2024

An Optimization Driven Link SINR Assurance in RIS-assisted Indoor Networks

Future smart factories are expected to deploy applications over high-performance indoor wireless channels in the millimeter-wave (mmWave) bands, which on the other hand are susceptible to high path losses and Line-of Sight (LoS) blockages. Low-cost Reconfigurable Intelligent Surfaces (RISs) can provide great opportunities in such scenarios, due to its ability to alleviate LoS link blockages. In this paper, we formulate a combinatorial optimization problem, solved with Integer Linear Programming (ILP) to optimally maintain connectivity by solving the problem of allocating RIS to robots in a wireless indoor network. Our model exploits the characteristic of nulling interference from RISs by tuning RIS reflection coefficients. We further consider Quality-of-Service (QoS) at receivers in terms of Signal-to-Interference-plus-Noise Ratio (SINR) and connection outages due to insufficient transmission quality service. Numerical results for optimal solutions and heuristics show the benefits of optimally deploying RISs by providing continuous connectivity through SINR, which significantly reduces outages due to link quality.

cs.NI↗

arXiv2024

SpiderDAN: Matching Augmentation in Demand-Aware Networks

Graph augmentation is a fundamental and well-studied problem that arises in network optimization. We consider a new variant of this model motivated by reconfigurable communication networks. In this variant, we consider a given physical network and the measured communication demands between the nodes. Our goal is to augment the given physical network with a matching, so that the shortest path lengths in the augmented network, weighted with the demands, are minimal.We prove that this problem is NP-hard, even if the physical network is a cycle. We then use results from demand-aware network design to provide a constant-factor approximation algorithm for adding a matching in case that only a few nodes in the network cause almost all the communication. For general real-world communication patterns, we design and evaluate a series of heuristics that can deal with arbitrary graphs as the underlying network structure. Our algorithms are validated experimentally using real-world traces (from e.g., Facebook) of data centers.

cs.DS↗

arXiv2024

Hash & Adjust: Competitive Demand-Aware Consistent Hashing

Distributed systems often serve dynamic workloads and resource demands evolve over time. Such a temporal behavior stands in contrast to the static and demand-oblivious nature of most data structures used by these systems. In this paper, we are particularly interested in consistent hashing, a fundamental building block in many large distributed systems. Our work is motivated by the hypothesis that a more adaptive approach to consistent hashing can leverage structure in the demand, and hence improve storage utilization and reduce access time. We initiate the study of demand-aware consistent hashing. Our main contribution is H&A, a constant-competitive online algorithm (i.e., it comes with provable performance guarantees over time). H&A is demand-aware and optimizes its internal structure to enable faster access times, while offering a high utilization of storage. We further evaluate H&A empirically.

cs.DC↗

arXiv2024

DecentPeeR: A Self-Incentivised & Inclusive Decentralized Peer Review System

Peer review, as a widely used practice to ensure the quality and integrity of publications, lacks a well-defined and common mechanism to self-incentivize virtuous behavior across all the conferences and journals. This is because information about reviewer efforts and author feedback typically remains local to a single venue, while the same group of authors and reviewers participate in the publication process across many venues. Previous attempts to incentivize the reviewing process assume that the quality of reviews and papers authored correlate for the same person, or they assume that the reviewers can receive physical rewards for their work. In this paper, we aim to keep track of reviewing and authoring efforts by users (who review and author) across different venues while ensuring self-incentivization. We show that our system, DecentPeeR, incentivizes reviewers to behave according to the rules, i.e., it has a unique Nash equilibrium in which virtuous behavior is rewarded.

cs.DL↗

arXiv2024

RIFO: Pushing the Efficiency of Programmable Packet Schedulers

Packet scheduling is a fundamental networking task that recently received renewed attention in the context of programmable data planes. Programmable packet scheduling systems such as those based on Push-In First-Out (PIFO) abstraction enabled flexible scheduling policies, but are too resource-expensive for large-scale line rate operation. This prompted research into practical programmable schedulers (e.g., SP-PIFO, AIFO) approximating PIFO behavior on regular hardware. Yet, their scalability remains limited due to extensive number of memory operations. To address this, we design an effective yet resource-efficient packet scheduler, Range-In First-Out (RIFO), which uses only three mutable memory cells and one FIFO queue per PIFO queue. RIFO is based on multi-criteria decision-making principles and uses small guaranteed admission buffers. Our large-scale simulations in Netbench demonstrate that despite using fewer resources, RIFO generally achieves competitive flow completion times across all studied workloads, and is especially effective in workloads with a significant share of large flows, reducing flow completion time up to 4.91x in datamining workload compared to state-of-the-art solutions. Our prototype implementation using P4 on Tofino switches requires only 650 lines of code, is scalable, and runs at line rate.

cs.NI↗

arXiv2024

Tight Bounds for Online Balanced Partitioning in the Generalized Learning Model

Resource allocation in distributed and networked systems such as the Cloud is becoming increasingly flexible, allowing these systems to dynamically adjust toward the workloads they serve, in a demand-aware manner. Online balanced partitioning is a fundamental optimization problem underlying such self-adjusting systems. We are given a set of $\ell$ servers. On each server we can schedule up to $k$ processes simultaneously. The demand is described as a sequence of requests $σ_t=\{p_i, p_{j}\}$, which means that the two processes $p_i,p_{j}$ communicate. A process can be migrated from one server to another which costs 1 unit per process move. If the communicating processes are on different servers, it further incurs a communication cost of 1 unit for this request. The objective is to minimize the competitive ratio: the cost of serving such a request sequence compared to the cost incurred by an optimal offline algorithm. Henzinger et al. (at SIGMETRICS 2019) introduced a learning variant of this problem where the cost of an online algorithm is compared to the cost of a static offline algorithm that does not perform any communication, but which simply learns the communication graph and keeps the discovered connected components together. This problem variant was recently also studied at SODA 2021. In this paper, we consider a more general learning model (i.e., stronger adversary), where the offline algorithm is not restricted to keep connected components together. Our main contribution are tight bounds for this problem. In particular, we present two deterministic online algorithms: (1) an online algorithm with competitive ratio $O(\max(\sqrt{k\ell \log k}, \ell \log k))$ and augmentation $1+ε$; (2) an online algorithm with competitive ratio $O(\sqrt{k})$ and augmentation $2+ε$. We further present lower bounds showing optimality of these bounds.

cs.DS↗

arXiv2024

On the Resilience of Fast Failover Routing Against Dynamic Link Failures

Modern communication networks feature local fast failover mechanisms in the data plane, to swiftly respond to link failures with pre-installed rerouting rules. This paper explores resilient routing meant to tolerate $\leq k$ simultaneous link failures, ensuring packet delivery, provided that the source and destination remain connected. While past theoretical works studied failover routing under static link failures, i.e., links which permanently and simultaneously fail, real-world networks often face link flapping--dynamic down states caused by, e.g., numerous short-lived software-related faults. Thus, in this initial work, we re-investigate the resilience of failover routing against link flapping, by categorizing link failures into static, semi-dynamic (removing the assumption that links fail simultaneously), and dynamic (removing the assumption that links fail permanently) types, shedding light on the capabilities and limitations of failover routing under these scenarios. We show that $k$-edge-connected graphs exhibit $(k-1)$-resilient routing against dynamic failures for $k \leq 5$. We further show that this result extends to arbitrary $k$ if it is possible to rewrite $\log k$ bits in the packet header. Rewriting $3$ bits suffices to cope with $k$ semi-dynamic failures. However, on general graphs, tolerating $2$ dynamic failures becomes impossible without bit-rewriting. Even by rewriting $\log k$ bits, resilient routing cannot resolve $k$ dynamic failures, demonstrating the limitation of local fast rerouting.

cs.NI↗

arXiv2024

Dynamic Probabilistic Reliable Broadcast

Byzantine reliable broadcast is a fundamental primitive in distributed systems that allows a set of processes to agree on a message broadcast by a dedicated process, even when some of them are malicious (Byzantine). It guarantees that no two correct processes deliver different messages, and if a message is delivered by a correct process, every correct process eventually delivers one. Byzantine reliable broadcast protocols are known to scale poorly, as they require $Ω(n^2)$ message exchanges, where $n$ is the number of system members. The quadratic cost can be explained by the inherent need for every process to relay a message to every other process. In this paper, we explore ways to overcome this limitation, by casting the problem to the probabilistic setting. We propose a solution in which every broadcast message is validated by a small set of witnesses, which allows us to maintain low latency and small communication complexity. In order to tolerate the slow adaptive adversary, we dynamically select the witnesses through a novel stream-local hash function: given a stream of inputs, it generates a stream of output hashed values that adapts to small deviations of the inputs. Our performance analysis shows that the proposed solution exhibits significant scalability gains over state-of-the-art protocols.

cs.DC↗

arXiv2024

Anomaly Detection Within Mission-Critical Call Processing

With increasingly larger and more complex telecommunication networks, there is a need for improved monitoring and reliability. Requirements increase further when working with mission-critical systems requiring stable operations to meet precise design and client requirements while maintaining high availability. This paper proposes a novel methodology for developing a machine learning model that can assist in maintaining availability (through anomaly detection) for client-server communications in mission-critical systems. To that end, we validate our methodology for training models based on data classified according to client performance. The proposed methodology evaluates the use of machine learning to perform anomaly detection of a single virtualized server loaded with simulated network traffic (using SIPp) with media calls. The collected data for the models are classified based on the round trip time performance experienced on the client side to determine if the trained models can detect anomalous client side performance only using key performance indicators available on the server. We compared the performance of seven different machine learning models by testing different trained and untrained test stressor scenarios. In the comparison, five models achieved an F1-score above 0.99 for the trained test scenarios. Random Forest was the only model able to attain an F1-score above 0.9 for all untrained test scenarios with the lowest being 0.980. The results suggest that it is possible to generate accurate anomaly detection to evaluate degraded client-side performance.

cs.NI↗

arXiv2024

Time Complexity of Broadcast and Consensus for Randomized Oblivious Message Adversaries

Broadcast and consensus are most fundamental tasks in distributed computing. These tasks are particularly challenging in dynamic networks where communication across the network links may be unreliable, e.g., due to mobility or failures. Indeed, over the last years, researchers have derived several impossibility results and high time complexity lower bounds (i.e., linear in the number of nodes $n$) for these tasks, even for oblivious message adversaries where communication networks are rooted trees. However, such deterministic adversarial models may be overly conservative, as many processes in real-world settings are stochastic in nature rather than worst case. This paper initiates the study of broadcast and consensus on stochastic dynamic networks, introducing a randomized oblivious message adversary. Our model is reminiscent of the SI model in epidemics, however, revolving around trees (which renders the analysis harder due to the apparent lack of independence). In particular, we show that if information dissemination occurs along random rooted trees, broadcast and consensus complete fast with high probability, namely in logarithmic time. Our analysis proves the independence of a key variable, which enables a formal understanding of the dissemination process. More formally, for a network with $n$ nodes, we first consider the completely random case where in each round the communication network is chosen uniformly at random among rooted trees. We then introduce the notion of randomized oblivious message adversary, where in each round, an adversary can choose $k$ edges to appear in the communication network, and then a rooted tree is chosen uniformly at random among the set of all rooted trees that include these edges. We show that broadcast completes in $O(k+\log n)$ rounds, and that this it is also the case for consensus as long as $k \le 0.1n$.

cs.DC↗

arXiv2024

A Demand-aware Networked System Using Telemetry and ML with ReactNET

Emerging network applications ranging from video streaming to virtual/augmented reality need to provide stringent quality-of-service (QoS) guarantees in complex and dynamic environments with shared resources. A promising approach to meeting these requirements is to automate complex network operations and create self-adjusting networks. These networks should automatically gather contextual information, analyze how to efficiently ensure QoS requirements, and adapt accordingly. This paper presents ReactNET, a self-adjusting networked system designed to achieve this vision by leveraging emerging network programmability and machine learning techniques. Programmability empowers ReactNET by providing fine-grained telemetry information, while machine learning-based classification techniques enable the system to learn and adjust the network to changing conditions. Our preliminary implementation of ReactNET in P4 and Python demonstrates its effectiveness in video streaming applications.

cs.NI↗

arXiv2024

Toward Self-Adjusting k-ary Search Tree Networks

Datacenter networks are becoming increasingly flexible with the incorporation of new networking technologies, such as optical circuit switches. These technologies allow for programmable network topologies that can be reconfigured to better serve network traffic, thus enabling a trade-off between the benefits (i.e., shorter routes) and costs of reconfigurations (i.e., overhead). Self-Adjusting Networks (SANs) aim at addressing this trade-off by exploiting patterns in network traffic, both when it is revealed piecewise (online dynamic topologies) or known in advance (offline static topologies). In this paper, we take the first steps toward Self-Adjusting k-ary tree networks. These are more powerful generalizations of existing binary search tree networks (like SplayNets), which have been at the core of SAN designs. k-ary search tree networks are a natural generalization offering nodes of higher degrees, reduced route lengths for a fixed number of nodes, and local routing in spite of reconfigurations. We first compute an offline (optimal) static network for arbitrary traffic patterns in $O(n^3 \cdot k)$ time via dynamic programming, and also improve the bound to $O(n^2 \cdot k)$ for the special case of uniformly distributed traffic. Then, we present a centroid-based topology of the network that can be used both in the offline static and the online setting. In the offline uniform-workload case, we construct this quasi-optimal network in linear time $O(n)$ and, finally, we present online self-adjusting k-ary search tree versions of SplayNet. We evaluate experimentally our new structure for $k=2$ (allowing for a comparison with existing SplayNets) on real and synthetic network traces. Our results show that this approach works better than SplayNet in most of the real network traces and in average to low locality synthetic traces, and is only little inferior to SplayNet in all remaining traces.

cs.NI↗

arXiv2024

D3: An Adaptive Reconfigurable Datacenter Network

The explosively growing communication traffic in datacenters imposes increasingly stringent performance requirements on the underlying networks. Over the last years, researchers have developed innovative optical switching technologies that enable reconfigurable datacenter networks (RCDNs) which support very fast topology reconfigurations. This paper presents D3, a novel and feasible RDCN architecture that improves throughput and flow completion time. D3 quickly and jointly adapts its links and packet scheduling toward the evolving demand, combining both demand-oblivious and demand-aware behaviors when needed. D3 relies on a decentralized network control plane supporting greedy, integrated-multihop, IP-based routing, allowing to react, quickly and locally, to topological changes without overheads. A rack-local synchronization and transport layer further support fast network adjustments. Moreover, we argue that D3 can be implemented using the recently proposed Sirius architecture (SIGCOMM 2020). We report on an extensive empirical evaluation using packet-level simulations. We find that D3 improves throughput by up to 15% and preserves competitive flow completion times compared to the state of the art. We further provide an analytical explanation of the superiority of D3, introducing an extension of the well-known Birkhoff-von Neumann decomposition, which may be of independent interest.

cs.NI↗