NetCruiser: Localize Network Failures by Learning from Latency Data

Data center networks (DCNs) are at the forefront of a radical transformation, fueled by the rapid rise of artificial intelligence (AI), cloud-native adoption, and enterprise digitalization, which together place unprecedented demands on their design and performance. From healthcare AI and intelligent transportation to business analytics, modern DCNs are expected to support a diverse and evolving mix of bandwidth-demanding applications, each with distinct quality of service (QoS) requirements. Yet, most state-of-the-art DCN solutions rely on oversimplified DCN data traffic models, failing to address the required application-level differentiation in terms of end-to-end (E2E) latency demands. In response, we propose ALAS-DCN, a novel, to our knowledge, application-aware, low-latency, scalable DCN architecture, based on optical packet switching in the intra-rack and inter-rack domains that is both designed and rigorously evaluated using commercial QoS specifications and realistic traffic patterns of different applications coexisting within the same DCN. ALAS-DCN integrates three hierarchical optical sub-networks: intra-rack, inter-rack, and bridge, each governed by a custom wavelength division multiplexing (WDM) medium access control (MAC) protocol that dynamically prioritizes critical traffic packets, according to the specific application E2E latency demands. Unlike prior approaches, our evaluation leverages realistic workload traces generated by established traffic generators, capturing the application diversity within modern data center (DC) environments. Extensive experimental simulations across five representative DC categories, each hosting a different mix of modern applications, demonstrate that ALAS-DCN consistently achieves up to 95% throughput efficiency and sub-µs latency for high priority traffic across a variety of applications, consistently meeting their differentiated latency constraints. Scalable up to 20,000 servers, ALAS-DCN redefines the DCN as not just a transport layer but an intelligent enabler of next-generation application performance.

A survey on architectures and energy efficiency in Data Center Networks

The vast majority of application traffic in modern data center networks (DCNs) can be classified into two categories: throughput-sensitive large flows and latency-sensitive small flows. These two types of flows have the conflicting requirements on link buffer occupancy. Existing data transfer proposals either do not fully utilize the path diversity of DCNs to improve the throughput of large flows, or cannot achieve a controllable link buffer occupancy to meet the low latency requirement of small flows. Aiming to balance throughput with latency, we develop the eXplicit MultiPath (XMP) congestion control scheme for DCNs. XMP comprises two components: the BOS algorithm brings link queue buffers consumed by large flows under control, while the TraSh algorithm is responsible for shifting traffic from more congested paths to less congested ones, thus achieving high throughput. We implemented XMP and evaluated its performance on traffic shifting, fairness, goodput, buffer occupancy and link utilization by conducting comprehensive experiments and simulations. The results show that XMP outperforms existing schemes and achieves a reasonable tradeoff between throughput and latency.

Many applications in public clouds require a high fan-in, many-to-one type of data communication (known as TCP incast) in modern Data Center Networks (DCNs). Such communication could cause severe incast congestion in switches and result in TCP throughput collapse, substantially degrading the application performance. The root cause of throughput collapse is the Retransmission Timeouts (RTO) due to packet losses in congested switches. Tenants in public clouds can opt to use a variety of TCP versions. However, the existing solutions rely on modifications of TCP protocols and specific techniques from switches, and thus these existing solutions are not always feasible for public clouds. In this paper, we are inspired by the emerging virtualization and network softwarization technologies to develop a novel scheme called Retransmission timeout Avoidance by Packet Intelligent Discarding (RAPID) using software switches. RAPID considers the number of packets of each incast flow, buffered in the switch to selectively discard some packets, and ensures that the Fast Retransmission/Fast Recovery rather than RTO is invoked at the sender(s) in response to packet loss. Thus, the long idle period of a timeout and the throughput drop are avoided. We prove that, given a predetermined minimum switch buffer space, dedicated to the incast application, RAPID can prevent RTO in all the incast senders. We also present a low-complexity heuristic version of RAPID named RAPID-ED, which combines the principles of RAPID and early detection and is extremely easy to implement on today's software switches. We evaluate the two proposed schemes in a data center network testbed built on NS-3 simulator. The simulation results confirm the theoretical expectation, and show that the RAPID and RAPID-ED perform very well to prevent RTO of TCP incast flows and hence the throughput collapse. Compared with other incast solutions, RAPID and RAPID-ED do not modify TCP protocols and therefore are more suitable in public clouds.

To meet the high performance requirement of real time and critical applications in industrial Internet of Things, modern lossless ethernet data center networks (DCNs) deployed with remote direct memory access and priority-based flow control (PFC) are dedicated to delivering low latency and high bandwidth. However, existing load balancing schemes either lack sub-round-trip-time congestion sensing or fail to accurately detect and reroute flows that cause congestion in PFC-enabled lossless DCNs. Therefore, we propose LBoDSN, an in-network load balancing for lossless DCNs using direct switch notification (DSN) for fast congestion flow detection, to address above challenges. LBoDSN tracks the evolution of ingress queue lengths at destination switches to anticipate the initiation of PFC pause, precisely identifies congested flows before PFC pause, and then sends DSNs to source switches to perform rerouting. After rerouting, the congestion notification packet associated with the previous path is selectively discarded to improve transmission performance. Experiments under realistic workloads reveal that, LBoDSN outperforms CONGA by 13%–65%, and 25%–80% in average and tail Flow Completion Times (FCTs), respectively. Compared to ConWeave, LBoDSN achieves approximately 9% improvement in both average and tail FCTs, while reducing switch queue consumption for reordering.

The Spanning Tree Protocol (STP), standardized in IEEE 802.1D, has served as the foundational loop-prevention mechanism in bridged Ethernet networks for over three decades. Despite its widespread adoption, STP exhibits significant limitations with respect to convergence time, bandwidth utilization, and scalability that are increasingly incompatible with the demands of modern enterprise and data center networks. This paper presents a structured comparative analysis of five major Layer 2 protocols in the STP lineage: STP (IEEE 802.1D), Rapid Spanning Tree Protocol (RSTP, IEEE 802.1w), Multiple Spanning Tree Protocol (MSTP, IEEE 802.1s), Shortest Path Bridging (SPB, IEEE 802.1aq), and Transparent Interconnection of Lots of Links (TRILL, RFC 6325). The analysis is conducted along four primary dimensions: convergence time, scalability, operational complexity, and security considerations. Drawing exclusively on peer-reviewed literature, IETF/IEEE standards, and published operational benchmarks, the paper concludes with a structured evaluation of the relevance of these protocols in Software-Defined Networking (SDN) environments and modern Ethernet fabric architectures, proposing a conceptual decision framework for protocol selection.

Middleboxes have become an integral part of modern enterprise and data center networks. They are used for realizing various performance and security objectives. Most middlboxes (e.g., firewalls, Intrusion Detection Systems (IDSs), Network Address Translators (NATs), etc.) are dedicated hardware appliances. However, recent advancements in cloud and virtualization technologies have fueled the concept of Virtual Middleboxes or Virtual Network Functions (VNFs) along with a new research field known as Network Function Virtualization (NFV). This area of research has gained a lot of traction from both industry and academia. Although much progress has been made in NFV technology, a crucial component for realizing the primary objective of NFV is still missing – a management and orchestration [7] system that conforms to the principles of NFV: open source, open API and standardized software solutions. Without this feature, network operators may end-up with the same situation of vendor lock-in as with proprietary hardware middleboxes. A number of recent proposals like Stratos [8], OpenNF [9], and Split/Merge [10], strive to fulfill the requirements for VNF management and orchestration. However, they propose incompatible northbound APIs. What is really needed is a standardized API that is flexible enough to express a wide range of NFV management and orchestration operations. History shows that standardization efforts usually take a long time and often are futile. Hence, we take a different approach, and propose to use an existing, well known, standardized interface for NFV management and orchestration: the Linux file system interface.

Along with the development of 5G and IoT technologies in recent years, Inter Data Center (Inter-DC) network is facing an explosive growth of geographically distributed user data, which needs to be duplicated among DCs in a real-time manner. Transmission-based applications require high availability that is going beyond 99.99%. However, with the expansion of Inter-DC network scale, link failures are also growing, which seriously affects data transmission efficiency, so fast link failure recovery is then urgently needed. Many previous works have been done to achieve fast failure recovery, but most of them ignore two key points, 1) the cost of deploying recovery strategies, and 2) the side-effect of re-transmission to network availability. These two factors make the existing failure recovery process too slow to be practical in real-time online industrial environments. To achieve realistic fast recovery from Inter-DC network failures, we propose a failure recovery framework FRAVaR, which achieves high network availability with very little deployment overhead. Particularly, FRAVaR reduces the deployment overhead by a novel incremental routing strategy to isolate link failures. In other words, it only needs to shuffle a tiny amount of traffic within a small failure isolation domain. On this base, FRAVaR further adopts a risk assessment theory named Value-at-Risk (VaR) to control flow re-transmission. We implement a prototype of FRAVaR and conduct a series of experiments on 4 real InterDC network topologies (ATT North America, IBM, GlobalCenter, AGIS). Experiment results show that FRAVaR outperforms state-of-the-art solutions on the recovery speed by 70.2%. <sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">1</sup>

Reducing traffic burstiness for MPTCP in data center networks

With the advent of software-defined networking (SDN) and OpenFlow switch protocol, routing and scheduling in modern data center networks is increasingly performed at the level of flows. A flow is a particular set of application traffic between two endpoints that receive the same forwarding decisions. As a consequence of the shift towards centralized flow-based control, efficient algorithms for scheduling and routing of flows have gained prominent importance. On the other hand, communication in datacenter applications, such as MAPREDUCE, and SPARK, often consists of multiple parallel data flows. Traditional methods to optimize flow metrics do not perform well since the flow abstraction is agnostic to application-level requirements, ignoring the fact that subsets of flows usually share the same performance goal. Chowdhury and Stoica introduced the concept ofcoflow as a networking abstraction that can capture the application-level semantics. In this dissertation, we design and analyze approximation algorithms for flow and coflow scheduling over various network models. In particular, we present (1) approximation algorithms for the problem of scheduling coflows over a general network topology with the goal of minimizing weighted average completion time; (2) improved algorithms for scheduling optical flows and coflows on a path; (3) a bi-criteria approximation algorithm for minimizing average response time of unit-size flows over a switch; (4) experimental work conducted to evaluate the practical performance of our algorithms and heuristics using our flow-level simulators.

Chapter 16 - Hypervisors, Virtualization, and Networking

Data storage optimizations (DS, e.g. low latency for data access) in data center networks(DCN) are difficult online-making problems. Previously, they are done with heuristics under static network models which highly rely on designers' understanding of the environment. Encouraged by recent successes in deep reinforcement learning techniques to solve intricate online assignment problems, we propose to use the Q-learning (QL) technique to train and learn from historical DS decisions, which can significantly reduce the data access delay. However, QL faces two challenges to be widely used in data centers. They are massive input data and the blindness on parameter settings which severely hamper the convergence of the learning process. To solve these two key problems, we develop an evolutionary QL scheme, named as LFDS (Low latency and Fast convergence Data Storage). In the initial stage of the LFDS, the input matrix of QL is sparse to shrink the dimensionality of the massive input data while retaining its information as much as possible. In the following training phase, a specialized neural network is adopted to achieves a quick approximation. To overcome the blindness during QL training, the two key parameters, learning rate, and discount rate are carefully tested with real data input and network architecture. The preferred range of learning rate and discount rate are recommended for the use of QL in data centers, which brings high training rewards and fast convergence. Extensive simulations with real-world data show that the data access latency is decreased by 23.5% and the convergence rate is increased by 15%.

Modern datacenter networks still operate through fragmented workflows in which predictive maintenance, intrusion detection, root cause analysis (RCA), and remediation are studied separately and deployed through loosely coupled tooling. This paper presents a unified AI system for autonomous self-healing datacenter networks that connects four stages: temporal failure prediction, drift-adaptive intrusion detection, topology-aware RCA, and safety-gated recovery with counterfactual validation. The architecture combines streaming telemetry, network-flow analytics, graph reasoning, and a topology digital twin inside a single operational loop. The system is formalized as a constrained sequential decision problem over telemetry, flows, topology, and policy constraints, and is evaluated through staged module validation plus a trace-driven closed-loop emulation. Because no public benchmark spans all four stages jointly, the empirical evidence combines public telemetry and flow datasets, streaming emulation, packet-capture replay, topology-grounded recovery traces, and a synthetic end-to-end incident timeline that makes the module handoff contract explicit. Across failure-prediction benchmarks, the temporal sequence model reaches F1 scores of 0.3737 on optical zero-shot hard-failure evaluation and 0.4677 on Cisco BGP failure prediction within a 60-second warning window. In intrusion detection, the drift-adaptive hybrid improves weighted F1 from 61.35% to 68.69% on full CICIDS2017 cross-dataset transfer without retraining the base detectors and reaches 98.05% weighted F1 in a packet-capture replay case study. For RCA, topology-aware reasoning reaches 0.8380 target-localization F1 with 1.0000 hidden-target accuracy and 0.9394 temporal RCA accuracy at 5.2 s mean detection delay. In the recovery twin, gated actions improve mean reachability from 0.9740 to 1.0000, achieve 0.8182 recovery success, and block 100% of mismatched unsafe actions. The results show that prediction, detection, diagnosis, and remediation can be organized into a reproducible closed loop for next-generation self-healing datacenter networks.

Modern datacenter networks are facing various challenges, e.g., highly dynamic workloads, congestion, topology asymmetry. ECMP, as a traditional load balancing mechanism which is widely used in today’s datacenters, can balance load poorly and lead to congestion. Variety of load balancing schemes are proposed to address the problems of ECMP. However, these traditional schemes usually make load balancing decision only based on network knowledge for a snapshot or a short time past. In this paper, we propose a Reinforcement Learning (RL) based approach, called RILNET (ReInforcement Learning NETworking), aiming at load balancing for datacenter networks. RILNET employs RL to learn a network and control it based on the learned experience. To achieve a higher granularity of control, RILNET is constructed to route flowlet rather than flows. Moreover, RILNET makes routing decisions for aggregation flows (an aggregation flow is a flow set that includes all flows flowing from the same source edge switch to the same destination edge switch) instead of a single flow. In order to test performance of RILNET, we propose a flow-level simulation and a packet-level simulation, and the both results show that RILNET can balance traffic load much more effectively than ECMP and another load balancing solution, i.e., DRILL. Compared with DRILL, RILNET outperforms DRILL in data loss and maximal link delay. Specifically, the maximal link data loss and the maximal link delay of RILNET are 44.4% and 25.4% smaller than DRILL, respectively.

Data center networks, using a reliable, high speed communication network, is a core component of cloud computing. Recently, optical networks using Space Division Multiplexing has been proposed as a solution to the speed bottleneck anticipated in data center networks. To our knowledge the new challenges of designing such communication systems have not been addressed yet. In this paper we have proposed an optimal approach to the problem of developing a path-protection scheme to handle communication requests in data center (DC) networks using elastic optical networking and space division multiplexing. We have formulated our problem as an Integer Linear Program (ILP). We have also proposed a heuristic that can handle problems of practical size. Our simulations explore important features of our approach.