TY - CONF AU - Dou, Feng AU - Wang, Lin AU - Chen, Shutong AU - Liu, Fangming ID - 50066 T2 - Proceedings of the IEEE International Conference on Computer Communications (INFOCOM) TI - X-Stream: A Flexible, Adaptive Video Transformer for Privacy-Preserving Video Stream Analytics ER - TY - CONF AU - Blöcher, Marcel AU - Nedderhut, Nils AU - Chuprikov, Pavel AU - Khalili, Ramin AU - Eugster, Patrick AU - Wang, Lin ID - 50065 T2 - Proceedings of the IEEE International Conference on Computer Communications (INFOCOM) TI - Train Once Apply Anywhere: Effective Scheduling for Network Function Chains Running on FUMES ER - TY - CONF AU - Hu, Haichuan AU - Liu, Fangming AU - Pei, Qiangyu AU - Yuan, Yongjie AU - Xu, Zichen AU - Wang, Lin ID - 50807 T2 - Proceedings of the ACM Web Conference (WWW) TI - 𝜆Grapher: A Resource-Efficient Serverless System for GNN Serving through Graph Sharing ER - TY - CONF AU - Razavi, Kamran AU - Ghafouri, Saeid AU - Mühlhäuser, Max AU - Jamshidi, Pooyan AU - Wang, Lin ID - 53095 T2 - Proceedings of the 4th Workshop on Machine Learning and Systems (EuroMLSys), colocated with EuroSys 2024 TI - Sponge: Inference Serving with Dynamic SLOs Using In-Place Vertical Scaling ER - TY - THES AU - Schneider, Stefan Balthasar ID - 29672 TI - Network and Service Coordination: Conventional and Machine Learning Approaches" ER - TY - CONF AB - Recent reinforcement learning approaches for continuous control in wireless mobile networks have shown impressive results. But due to the lack of open and compatible simulators, authors typically create their own simulation environments for training and evaluation. This is cumbersome and time-consuming for authors and limits reproducibility and comparability, ultimately impeding progress in the field. To this end, we propose mobile-env, a simple and open platform for training, evaluating, and comparing reinforcement learning and conventional approaches for continuous control in mobile wireless networks. mobile-env is lightweight and implements the common OpenAI Gym interface and additional wrappers, which allows connecting virtually any single-agent or multi-agent reinforcement learning framework to the environment. While mobile-env provides sensible default values and can be used out of the box, it also has many configuration options and is easy to extend. We therefore believe mobile-env to be a valuable platform for driving meaningful progress in autonomous coordination of wireless mobile networks. AU - Schneider, Stefan Balthasar AU - Werner, Stefan AU - Khalili, Ramin AU - Hecker, Artur AU - Karl, Holger ID - 30236 KW - wireless mobile networks KW - network management KW - continuous control KW - cognitive networks KW - autonomous coordination KW - reinforcement learning KW - gym environment KW - simulation KW - open source T2 - IEEE/IFIP Network Operations and Management Symposium (NOMS) TI - mobile-env: An Open Platform for Reinforcement Learning in Wireless Mobile Networks ER - TY - CONF AB - The decentralized nature of multi-agent systems requires continuous data exchange to achieve global objectives. In such scenarios, Age of Information (AoI) has become an important metric of the freshness of exchanged data due to the error-proneness and delays of communication systems. Communication systems usually possess dependencies: the process describing the success or failure of communication is highly correlated when these attempts are ``close'' in some domain (e.g. in time, frequency, space or code as in wireless communication) and is, in general, non-stationary. To study AoI in such scenarios, we consider an abstract event-based AoI process $\Delta(n)$, expressing time since the last update: If, at time $n$, a monitoring node receives a status update from a source node (event $A(n-1)$ occurs), then $\Delta(n)$ is reset to one; otherwise, $\Delta(n)$ grows linearly in time. This AoI process can thus be viewed as a special random walk with resets. The event process $A(n)$ may be nonstationary and we merely assume that its temporal dependencies decay sufficiently, described by $\alpha$-mixing. We calculate moment bounds for the resulting AoI process as a function of the mixing rate of $A(n)$. Furthermore, we prove that the AoI process $\Delta(n)$ is itself $\alpha$-mixing from which we conclude a strong law of large numbers for $\Delta(n)$. These results are new, since AoI processes have not been studied so far in this general strongly mixing setting. This opens up future work on renewal processes with non-independent interarrival times. AU - Redder, Adrian AU - Ramaswamy, Arunselvan AU - Karl, Holger ID - 32811 T2 - Proceedings of the 58th Allerton Conference on Communication, Control, and Computing TI - Age of Information Process under Strongly Mixing Communication -- Moment Bound, Mixing Rate and Strong Law ER - TY - CONF AU - Redder, Adrian AU - Ramaswamy, Arunselvan AU - Karl, Holger ID - 30793 T2 - Proceedings of the 14th International Conference on Agents and Artificial Intelligence TI - Multi-agent Policy Gradient Algorithms for Cyber-physical Systems with Lossy Communication ER - TY - GEN AB - Iterative distributed optimization algorithms involve multiple agents that communicate with each other, over time, in order to minimize/maximize a global objective. In the presence of unreliable communication networks, the Age-of-Information (AoI), which measures the freshness of data received, may be large and hence hinder algorithmic convergence. In this paper, we study the convergence of general distributed gradient-based optimization algorithms in the presence of communication that neither happens periodically nor at stochastically independent points in time. We show that convergence is guaranteed provided the random variables associated with the AoI processes are stochastically dominated by a random variable with finite first moment. This improves on previous requirements of boundedness of more than the first moment. We then introduce stochastically strongly connected (SSC) networks, a new stochastic form of strong connectedness for time-varying networks. We show: If for any $p \ge0$ the processes that describe the success of communication between agents in a SSC network are $\alpha$-mixing with $n^{p-1}\alpha(n)$ summable, then the associated AoI processes are stochastically dominated by a random variable with finite $p$-th moment. In combination with our first contribution, this implies that distributed stochastic gradient descend converges in the presence of AoI, if $\alpha(n)$ is summable. AU - Redder, Adrian AU - Ramaswamy, Arunselvan AU - Karl, Holger ID - 30790 T2 - arXiv:2201.11343 TI - Distributed gradient-based optimization in the presence of dependent aperiodic communication ER - TY - GEN AB - We present sufficient conditions that ensure convergence of the multi-agent Deep Deterministic Policy Gradient (DDPG) algorithm. It is an example of one of the most popular paradigms of Deep Reinforcement Learning (DeepRL) for tackling continuous action spaces: the actor-critic paradigm. In the setting considered herein, each agent observes a part of the global state space in order to take local actions, for which it receives local rewards. For every agent, DDPG trains a local actor (policy) and a local critic (Q-function). The analysis shows that multi-agent DDPG using neural networks to approximate the local policies and critics converge to limits with the following properties: The critic limits minimize the average squared Bellman loss; the actor limits parameterize a policy that maximizes the local critic's approximation of $Q_i^*$, where $i$ is the agent index. The averaging is with respect to a probability distribution over the global state-action space. It captures the asymptotics of all local training processes. Finally, we extend the analysis to a fully decentralized setting where agents communicate over a wireless network prone to delays and losses; a typical scenario in, e.g., robotic applications. AU - Redder, Adrian AU - Ramaswamy, Arunselvan AU - Karl, Holger ID - 30791 T2 - arXiv:2201.00570 TI - Asymptotic Convergence of Deep Multi-Agent Actor-Critic Algorithms ER - TY - JOUR AU - Redder, Adrian AU - Ramaswamy, Arunselvan AU - Karl, Holger ID - 32854 IS - 13 JF - IFAC-PapersOnLine TI - Practical Network Conditions for the Convergence of Distributed Optimization VL - 55 ER - TY - CONF AB - Modern services often comprise several components, such as chained virtual network functions, microservices, or machine learning functions. Providing such services requires to decide how often to instantiate each component, where to place these instances in the network, how to chain them and route traffic through them. To overcome limitations of conventional, hardwired heuristics, deep reinforcement learning (DRL) approaches for self-learning network and service management have emerged recently. These model-free DRL approaches are more flexible but typically learn tabula rasa, i.e., disregard existing understanding of networks, services, and their coordination. Instead, we propose FutureCoord, a novel model-based AI approach that leverages existing understanding of networks and services for more efficient and effective coordination without time-intensive training. FutureCoord combines Monte Carlo Tree Search with a stochastic traffic model. This allows FutureCoord to estimate the impact of future incoming traffic and effectively optimize long-term effects, taking fluctuating demand and Quality of Service (QoS) requirements into account. Our extensive evaluation based on real-world network topologies, services, and traffic traces indicates that FutureCoord clearly outperforms state-of-the-art model-free and model-based approaches with up to 51% higher flow success ratios. AU - Werner, Stefan AU - Schneider, Stefan Balthasar AU - Karl, Holger ID - 29220 KW - network management KW - service management KW - AI KW - Monte Carlo Tree Search KW - model-based KW - QoS T2 - IEEE/IFIP Network Operations and Management Symposium (NOMS) TI - Use What You Know: Network and Service Coordination Beyond Certainty ER - TY - CONF AB - Datacenter applications have different resource requirements from network and developing flow scheduling heuristics for every workload is practically infeasible. In this paper, we show that deep reinforcement learning (RL) can be used to efficiently learn flow scheduling policies for different workloads without manual feature engineering. Specifically, we present LFS, which learns to optimize a high-level performance objective, e.g., maximize the number of flow admissions while meeting the deadlines. The LFS scheduler is trained through deep RL to learn a scheduling policy on continuous online flow arrivals. The evaluation results show that the trained LFS scheduler admits 1.05x more flows than the greedy flow scheduling heuristics under varying network load. AU - Hasnain, Asif AU - Karl, Holger ID - 20125 KW - Flow scheduling KW - Deadlines KW - Reinforcement learning T2 - 2021 IEEE 18th Annual Consumer Communications & Networking Conference (CCNC) TI - Learning Flow Scheduling ER - TY - THES AU - Hasnain, Asif ID - 27503 TI - Automating Network Resource Allocation for Coflows with Deadlines ER - TY - CONF AB - Data-parallel applications are developed using different data programming models, e.g., MapReduce, partition/aggregate. These models represent diverse resource requirements of application in a datacenter network, which can be represented by the coflow abstraction. The conventional method of creating hand-crafted coflow heuristics for admission or scheduling for different workloads is practically infeasible. In this paper, we propose a deep reinforcement learning (DRL)-based coflow admission scheme -- LCS -- that can learn an admission policy for a higher-level performance objective, i.e., maximize successful coflow admissions, without manual feature engineering. LCS is trained on a production trace, which has online coflow arrivals. The evaluation results show that LCS is able to learn a reasonable admission policy that admits more coflows than state-of-the-art Varys heuristic while meeting their deadlines. AU - Hasnain, Asif AU - Karl, Holger ID - 21005 KW - Coflow scheduling KW - Reinforcement learning KW - Deadlines T2 - IEEE INFOCOM 2021 - IEEE Conference on Computer Communications Workshops (INFOCOM WKSHPS) TI - Learning Coflow Admissions ER - TY - CONF AB - Services often consist of multiple chained components such as microservices in a service mesh, or machine learning functions in a pipeline. Providing these services requires online coordination including scaling the service, placing instance of all components in the network, scheduling traffic to these instances, and routing traffic through the network. Optimized service coordination is still a hard problem due to many influencing factors such as rapidly arriving user demands and limited node and link capacity. Existing approaches to solve the problem are often built on rigid models and assumptions, tailored to specific scenarios. If the scenario changes and the assumptions no longer hold, they easily break and require manual adjustments by experts. Novel self-learning approaches using deep reinforcement learning (DRL) are promising but still have limitations as they only address simplified versions of the problem and are typically centralized and thus do not scale to practical large-scale networks. To address these issues, we propose a distributed self-learning service coordination approach using DRL. After centralized training, we deploy a distributed DRL agent at each node in the network, making fast coordination decisions locally in parallel with the other nodes. Each agent only observes its direct neighbors and does not need global knowledge. Hence, our approach scales independently from the size of the network. In our extensive evaluation using real-world network topologies and traffic traces, we show that our proposed approach outperforms a state-of-the-art conventional heuristic as well as a centralized DRL approach (60% higher throughput on average) while requiring less time per online decision (1 ms). AU - Schneider, Stefan Balthasar AU - Qarawlus, Haydar AU - Karl, Holger ID - 21543 KW - network management KW - service management KW - coordination KW - reinforcement learning KW - distributed T2 - IEEE International Conference on Distributed Computing Systems (ICDCS) TI - Distributed Online Service Coordination Using Deep Reinforcement Learning ER - TY - CONF AB - In practical, large-scale networks, services are requested by users across the globe, e.g., for video streaming. Services consist of multiple interconnected components such as microservices in a service mesh. Coordinating these services requires scaling them according to continuously changing user demand, deploying instances at the edge close to their users, and routing traffic efficiently between users and connected instances. Network and service coordination is commonly addressed through centralized approaches, where a single coordinator knows everything and coordinates the entire network globally. While such centralized approaches can reach global optima, they do not scale to large, realistic networks. In contrast, distributed approaches scale well, but sacrifice solution quality due to their limited scope of knowledge and coordination decisions. To this end, we propose a hierarchical coordination approach that combines the good solution quality of centralized approaches with the scalability of distributed approaches. In doing so, we divide the network into multiple hierarchical domains and optimize coordination in a top-down manner. We compare our hierarchical with a centralized approach in an extensive evaluation on a real-world network topology. Our results indicate that hierarchical coordination can find close-to-optimal solutions in a fraction of the runtime of centralized approaches. AU - Schneider, Stefan Balthasar AU - Jürgens, Mirko AU - Karl, Holger ID - 20693 KW - network management KW - service management KW - coordination KW - hierarchical KW - scalability KW - nfv T2 - IFIP/IEEE International Symposium on Integrated Network Management (IM) TI - Divide and Conquer: Hierarchical Network and Service Coordination ER - TY - JOUR AB - Modern services consist of interconnected components,e.g., microservices in a service mesh or machine learning functions in a pipeline. These services can scale and run across multiple network nodes on demand. To process incoming traffic, service components have to be instantiated and traffic assigned to these instances, taking capacities, changing demands, and Quality of Service (QoS) requirements into account. This challenge is usually solved with custom approaches designed by experts. While this typically works well for the considered scenario, the models often rely on unrealistic assumptions or on knowledge that is not available in practice (e.g., a priori knowledge). We propose DeepCoord, a novel deep reinforcement learning approach that learns how to best coordinate services and is geared towards realistic assumptions. It interacts with the network and relies on available, possibly delayed monitoring information. Rather than defining a complex model or an algorithm on how to achieve an objective, our model-free approach adapts to various objectives and traffic patterns. An agent is trained offline without expert knowledge and then applied online with minimal overhead. Compared to a state-of-the-art heuristic, DeepCoord significantly improves flow throughput (up to 76%) and overall network utility (more than 2x) on realworld network topologies and traffic traces. It also supports optimizing multiple, possibly competing objectives, learns to respect QoS requirements, generalizes to scenarios with unseen, stochastic traffic, and scales to large real-world networks. For reproducibility and reuse, our code is publicly available. AU - Schneider, Stefan Balthasar AU - Khalili, Ramin AU - Manzoor, Adnan AU - Qarawlus, Haydar AU - Schellenberg, Rafael AU - Karl, Holger AU - Hecker, Artur ID - 21808 JF - Transactions on Network and Service Management KW - network management KW - service management KW - coordination KW - reinforcement learning KW - self-learning KW - self-adaptation KW - multi-objective TI - Self-Learning Multi-Objective Service Coordination Using Deep Reinforcement Learning ER - TY - GEN AB - Macrodiversity is a key technique to increase the capacity of mobile networks. It can be realized using coordinated multipoint (CoMP), simultaneously connecting users to multiple overlapping cells. Selecting which users to serve by how many and which cells is NP-hard but needs to happen continuously in real time as users move and channel state changes. Existing approaches often require strict assumptions about or perfect knowledge of the underlying radio system, its resource allocation scheme, or user movements, none of which is readily available in practice. Instead, we propose three novel self-learning and self-adapting approaches using model-free deep reinforcement learning (DRL): DeepCoMP, DD-CoMP, and D3-CoMP. DeepCoMP leverages central observations and control of all users to select cells almost optimally. DD-CoMP and D3-CoMP use multi-agent DRL, which allows distributed, robust, and highly scalable coordination. All three approaches learn from experience and self-adapt to varying scenarios, reaching 2x higher Quality of Experience than other approaches. They have very few built-in assumptions and do not need prior system knowledge, making them more robust to change and better applicable in practice than existing approaches. AU - Schneider, Stefan Balthasar AU - Karl, Holger AU - Khalili, Ramin AU - Hecker, Artur ID - 33854 KW - mobility management KW - coordinated multipoint KW - CoMP KW - cell selection KW - resource management KW - reinforcement learning KW - multi agent KW - MARL KW - self-learning KW - self-adaptation KW - QoE TI - DeepCoMP: Coordinated Multipoint Using Multi-Agent Deep Reinforcement Learning ER - TY - GEN AB - Network and service coordination is important to provide modern services consisting of multiple interconnected components, e.g., in 5G, network function virtualization (NFV), or cloud and edge computing. In this paper, I outline my dissertation research, which proposes six approaches to automate such network and service coordination. All approaches dynamically react to the current demand and optimize coordination for high service quality and low costs. The approaches range from centralized to distributed methods and from conventional heuristic algorithms and mixed-integer linear programs to machine learning approaches using supervised and reinforcement learning. I briefly discuss their main ideas and advantages over other state-of-the-art approaches and compare strengths and weaknesses. AU - Schneider, Stefan Balthasar ID - 35889 KW - nfv KW - coordination KW - machine learning KW - reinforcement learning KW - phd KW - digest TI - Conventional and Machine Learning Approaches for Network and Service Coordination ER -