PaperlayerView exact revision on arXiv

Source-linked AI summary

Low-latency Networking: Where Latency Lurks and How to Tame It

Xiaolin Jiang, Hossein S. Ghadikolaei, Gabor Fodor, Eytan Modiano, Zhibo Pang, Michele Zorzi, Carlo Fischione

arXiv:1808.02079v1eess.SPcs.NI

TL;DR

IoT and mission-critical services require millisecond or sub-millisecond communication, but latency emerges from complex interactions across network layers and domains. The paper provides a holistic analysis and classification of design principles and enabling technologies, concluding that they must be jointly optimized to reduce delay while satisfying reliability and other performance requirements.

  • Problem

    Millisecond and sub-millisecond services expose complex cross-layer and multi-domain latency interactions that existing communication-network designs do not straightforwardly address.

  • Method

    The paper holistically analyzes and classifies latency components, design principles, enabling technologies, standardization activities, and open research problems.

  • Results

    The analysis shows that low-latency techniques affect one or multiple delay components and must be optimized together with other requirements.

  • Takeaways & Limitations

    Deploying low-latency networks requires coordinating enabling technologies across layers and managing trade-offs among latency, reliability, complexity, and processing delay.

Abstract

from arXiv · show

While the current generation of mobile and fixed communication networks has been standardized for mobile broadband services, the next generation is driven by the vision of the Internet of Things and mission critical communication services requiring latency in the order of milliseconds or sub-milliseconds. However, these new stringent requirements have a large technical impact on the design of all layers of the communication protocol stack. The cross layer interactions are complex due to the multiple design principles and technologies that contribute to the layers' design and fundamental performance limitations. We will be able to develop low-latency networks only if we address the problem of these complex interactions from the new point of view of sub-milliseconds latency. In this article, we propose a holistic analysis and classification of the main design principles and enabling technologies that will make it possible to deploy low-latency wireless communication networks. We argue that these design principles and enabling technologies must be carefully orchestrated to meet the stringent requirements and to manage the inherent trade-offs between low latency and traditional performance metrics. We also review currently ongoing standardization activities in prominent standards associations, and discuss open problems for future research.

I. INTRODUCTION

Next-generation networks target millisecond or sub-millisecond latency for IoT and mission-critical services, but achieving it requires coordinated changes across complex, multi-domain protocol-stack interactions.

  • I. INTRODUCTION: Millisecond and sub-millisecond services motivate the next communication-network revolution, including IoT, URLLC, and mission-critical applications.The cited use cases include industrial, healthcare, transportation, smart-grid, and manufacturing scenarios.
  • I. INTRODUCTION: Latency is difficult to control because end-to-end routes span multiple wireless, local, core, and service-network domains with interacting traffic, access, and processing factors.The paper identifies link sharing, medium access, networking technologies, competing traffic, and service-processing algorithms as latency factors.
  • I. INTRODUCTION: The paper analyzes and classifies design principles and enabling technologies across physical, medium-access, network, and transport layers to support networks below 1 ms latency.The analysis covers access, core, and service networks and considers both latency and reliability requirements.
  • I. INTRODUCTION: Low-latency techniques must be orchestrated across layers while balancing latency against reliability and other network-performance requirements.Different use cases impose different reliability levels, which can change the available techniques for achieving a given latency target.
  • I. INTRODUCTION: The article structures its analysis around application requirements, latency components, single-domain technologies, and inter-domain latency reduction and control.The paper examines both proximal and long-distance communication, including multi-domain networks.

III. CAUSES AND DEFINITIONS OF LATENCY COMPONENTS IN SINGLE HOP, MULTI-HOP AND MULTI-DOMAIN NETWORKS

The paper defines end-to-end latency as packet generation to correct reception and decomposes it into queueing, processing, access, transmission, and routing components across network paths.

  • III. CAUSES AND DEFINITIONS OF LATENCY COMPONENTS IN SINGLE HOP, MULTI-HOP AND MULTI-DOMAIN NETWORKS: End-to-end latency is the time from packet generation to correct reception, but cross-domain guarantees are difficult because operators lack control over all contributing networks.Different domains may introduce random latencies whose exact values or distributions are unknown in advance.
  • III. CAUSES AND DEFINITIONS OF LATENCY COMPONENTS IN SINGLE HOP, MULTI-HOP AND MULTI-DOMAIN NETWORKS: Latency accumulates from queueing and processing, medium access, transmission, and routing components.The paper denotes these components as Tq + Tpr, Ta, Tt, and Tr, respectively.
  • III. CAUSES AND DEFINITIONS OF LATENCY COMPONENTS IN SINGLE HOP, MULTI-HOP AND MULTI-DOMAIN NETWORKS: Single-hop latency includes waiting and processing plus access to a shared medium, including control signaling such as RTS/CTS or scheduling grants.These access procedures contribute to Ta even before packet transmission completes.
  • III. CAUSES AND DEFINITIONS OF LATENCY COMPONENTS IN SINGLE HOP, MULTI-HOP AND MULTI-DOMAIN NETWORKS: Multi-hop paths add the access and transmission times of each hop and may add routing time for selecting the outgoing interface.Routing latency can be zero when forwarding occurs at the MAC layer without routing.
  • III. CAUSES AND DEFINITIONS OF LATENCY COMPONENTS IN SINGLE HOP, MULTI-HOP AND MULTI-DOMAIN NETWORKS: Random interference, shadowing, fading, and fluctuating traffic loads make end-to-end latency difficult to predict and control.A tagged stream is affected by simultaneously delivered traffic and changing wireless-channel conditions.

IV. INTRA-NETWORK TECHNIQUES AND TECHNOLOGIES

Low-latency network design requires reducing specific or combined latency components through cross-layer techniques while preserving spectral efficiency, energy efficiency, data rate, and capacity.

  • IV. INTRA-NETWORK TECHNIQUES AND TECHNOLOGIES: Cross-layer techniques target queueing, processing, access, transmission, and routing components individually or in combination.The paper emphasizes that physical, MAC, network, and transport layers must coordinate their effects on latency.
  • IV. INTRA-NETWORK TECHNIQUES AND TECHNOLOGIES: Candidate technologies must reduce latency while accounting for spectral efficiency, energy efficiency, peak data rate, and capacity constraints.The discussion also includes technologies not specifically designed for latency-critical services but potentially useful for meeting latency targets.

A. Physical Layer Techniques

Physical-layer choices reduce transmission time through higher rates, adaptation, bandwidth, and waveform design, but they trade latency against reliability, synchronization, complexity, and processing delay.

  • A. Physical Layer Techniques: Higher transmission rates reduce fixed-packet transmission time, while PHY design must balance latency, reliability, spectral efficiency, energy efficiency, and range.The maximum rate depends on bandwidth and instantaneous SINR.
  • A. Physical Layer Techniques: Adaptive modulation and coding raises transmission rate at high SINR but lowers the rate under poor channels or high interference to maintain target BER.Adaptation can use modulation order, coding strength, transmit power, or combinations of these parameters.
  • A. Physical Layer Techniques: Increasing bandwidth can raise transmission rate and reduce latency before capacity gains become negligible, although bandwidth also changes SINR.Capacity is therefore not a linear function of bandwidth.
  • 2) Waveform Design:: Traditional OFDM can impose processing and synchronization costs that are problematic for short packets and stringent latency requirements.OFDM requires precise synchronization to preserve sub-carrier orthogonality and avoid inter-carrier interference.
  • 2) Waveform Design:: Asynchronous waveforms such as GFDM relax synchronization requirements and may reduce transmission components, but filtering and processing can increase complexity and latency.GFDM can use flexible symbol durations and a single cyclic prefix, while high transmitter filter order increases processing latency.
  • A. Physical Layer Techniques: Short packets reduce transmission time but reduce the maximum coding rate for a fixed packet-error probability by a factor proportional to 1/√n.Finite blocklength therefore creates a direct latency–coding-rate trade-off.

3) Sending Short Packets:

Short packets make metadata overhead significant, while MAC functions and initial access can determine effective latency beyond the physical transmission rate.

  • 3) Sending Short Packets:: Short packets make low-rate metadata coding inefficient because metadata size becomes comparable to payload size.For long packets, metadata overhead is negligible; for short packets, it cannot be ignored.
  • 3) Sending Short Packets:: MAC-layer choices can substantially reduce effective transmission rate through inefficient access, queue management, and scheduling.The MAC layer handles synchronization, initial access, interference management, scheduling, and rate adaptation.
  • 3) Sending Short Packets:: One-hop access latency includes processing, queuing, association, delay-to-access, and transmission-attempt components.Processing and queuing latencies are included in the definition but are not covered in the paper.
  • 3) Sending Short Packets:: Initial access and association define how a new device connects to the network and can become important for short, sporadic IoT transmissions.These functions are introduced as core MAC-layer procedures for new devices.

1) Initial Access and Association:

Initial access and MAC scheduling must balance latency guarantees against signaling, computation, collisions, and scalability for massive low-payload device populations.

  • 1) Initial Access and Association:: For massive wake-up radios sending only a few bits, association latency can become problematic despite being tolerable for continuously connected or large-data devices.The relevant boundary is the combination of many devices, infrequent wake-up, and very small payloads.
  • 1) Initial Access and Association:: Contention-free MAC offers bounded latency and jitter but incurs signaling and computational overhead that may not scale to massive short-packet access.With TDMA for N devices, average latency is N/2×Tt, while other schemes still require access-request registration and instructions.
  • 1) Initial Access and Association:: Contention-based MAC reduces signaling overhead but introduces random channel-establishment latency and collision-dependent retransmissions.Slotted-ALOHA performs well at low offered load but degrades sharply as repeated collisions increase backoff.
  • 1) Initial Access and Association:: CSMA improves high-load performance by sensing the channel, although hidden and exposed nodes can still cause collisions or unnecessary deferral.Its channel assessment and random backoff reduce collision probability relative to simpler contention-based access.
  • 1) Initial Access and Association:: With directional communication, slotted-ALOHA and CSMA can outperform TDMA in throughput and delay, reducing the need for complex scheduling when beamforming is timely.The cited analysis identifies directionality as important for low-latency massive access.
  • 1) Initial Access and Association:: Correlatable symbol sequences replace traditional control messages with rapidly decodable signaling for distributed channel reservation.The technique is proposed in 802.11ec and extended to mmWave networks to address control/data-plane rate imbalance.

3) Mobility Management:

Mobility and routing mechanisms must reduce handover and forwarding delay while coping with directional links, dynamic topologies, interference, and throughput trade-offs.

  • 3) Mobility Management:: Location-only handover prediction is limited in directional networks because fixed terminals may also need handover when obstacles block links.Frequent handovers in mmWave networks motivate soft handover and coordination strategies.
  • 3) Mobility Management:: Long-lived sensors require an energy-delay-spectral-efficiency trade-off because they tolerate hundreds of milliseconds but need years or decades of battery life.Latency requirements are therefore not uniform across applications.
  • 3) Mobility Management:: Throughput-optimal back-pressure routing can incur quadratic end-to-end queuing delay with hop count, motivating variants that trade minor throughput loss for linear delay growth.Shadow queues are used to reduce the number of real queues in the modified algorithm.
  • 3) Mobility Management:: Overlay routing lets a subset of nodes run back-pressure while legacy nodes use shortest-path routing, achieving full throughput with a small overlay in simulation.This addresses the difficulty of upgrading every device in the network.
  • 3) Mobility Management:: Wireless interference strongly affects scheduling delay, and delay-oriented back-pressure variants can approach a derived lower bound while optimizing energy and spectral-efficiency trade-offs.The lower bound applies to multihop wireless networks with fixed routes.
  • 3) Mobility Management:: Queue-length routing metrics can delay light-traffic flows, motivating head-of-line delay and Q-metric alternatives.These alternatives target cases where short queues receive little service under queue-length prioritization.

2) Non-back-pressure-based Routing:

Non-back-pressure routing spans source-routed, geographic, proactive, and ad hoc approaches, exposing fundamental throughput-delay trade-offs and topology-discovery costs.

  • 2) Non-back-pressure-based Routing:: Universal Max-Weight uses source routing and virtual-network backlog costs to avoid back-pressure loops and significantly reduce delay relative to back-pressure variants.Packets follow selected shortest paths under the source-routing formulation.
  • 2) Non-back-pressure-based Routing:: Dynamic ad hoc networks make topology discovery non-negligible, so routing methods designed for fixed known topologies may not meet low-latency needs.The section therefore turns to routing classes suited to unknown and changing topologies.
  • 2) Non-back-pressure-based Routing:: In static random networks, optimal per-node throughput scales as Θ(1/√n log n) while delay scales as Θ(n/√log n), revealing a throughput-delay trade-off.Transmission power can reduce delay through longer range but increases interference and can reduce throughput.
  • 2) Non-back-pressure-based Routing:: Mobility can raise achievable throughput to Θ(1) but increases delay scaling to Θ(n log n), illustrating a direct capacity-delay trade-off.The cited result does not include channel-estimation, signaling, or computational overheads, so actual delay may be larger.
  • 2) Non-back-pressure-based Routing:: Multicast redundancy lowers two-hop relay delay but decreases capacity, while dense-network broadcast scales with capacity 1/n and delay log log n.The cited broadcast results report worse throughput and delay scaling in sparse networks.
  • 2) Non-back-pressure-based Routing:: SPEED provides per-hop delay guarantees and bounds hop count using localized geographic forwarding with compensation for local decision errors.Its extension supports multiple QoS levels measured by reliability and latency.
  • 2) Non-back-pressure-based Routing:: Proactive distance-vector routing can guarantee end-to-end latency by probing shortest paths for suitability before delay-sensitive transmission.Nodes maintain distance and next-hop tables for every destination.

E. Transmission Capacity Boosting and Sharing Techniques

Capacity-oriented wireless techniques can reduce low-latency communication delays, but their gains introduce access, coordination, or processing costs that must be managed.

  • 1) MU-MIMO:: MU-MIMO serves multiple users in the same time-frequency channel, reducing the MAC-layer delay-to-access component.Beamforming steers energy toward intended locations while minimizing interference.
  • 1) MU-MIMO:: Large antenna arrays simplify beamforming asymptotically, but CSI acquisition delay can limit their suitability for low-latency services.The limitation is exacerbated in mmWave communications, and low-latency MIMO beamforming remains largely open.
  • 2) Millimeter-wave:: mmWave networks achieve almost negligible transmission latency, but pilot transmission and beamforming design can create delay-to-access.Large bandwidth and antenna arrays support extremely high data rates while reducing interference.
  • 2) Millimeter-wave:: Digital beamforming with low-resolution A/D converters and flexible transmission time intervals were reported as suitable for reducing beamforming and MAC-layer delay.The study targeted 20 Gbps data rate and 1 ms delay requirements in mmWave cellular networks.

3) Full-duplex:

Low-latency wireless design combines full-duplex, direct device communication, fronthaul architectures, and retransmission schemes, each trading latency gains against interference, processing, or reliability constraints.

  • 3) Full-duplex:: Self-interference suppression enables full-duplex gains, but active suppression increases processing delay.Passive suppression primarily concerns antenna design, while active suppression uses signal processing and beamforming.
  • 3) Full-duplex:: Full duplex reduces transmission, MAC, and routing latency by enabling simultaneous transmission and reception on shared resources.Its benefits depend on effective cancellation and suitable MAC-layer design.
  • 4) Device-to-Device Communication:: D2D communication reduces routing delay by using single-hop or fewer-hop paths and separating local traffic from cellular core-network traffic.Shared-resource D2D requires allocation mechanisms to control interference and preserve latency for both traffic types.
  • 5) Cloud Radio Access Network and Mobile Fronthaul:: Cloud RAN and mobile fronthaul can reduce routing and processing latency through centralized processing, joint communication, and reconfigurable optical topologies.A fiber–mmWave combination supports mobility and reduces optic-to-mmWave conversion latency through analog waveform transmission.
  • 6) Forward Error Correction and Hybrid Automatic Repeat Request Techniques:: Block acknowledgements reduce MAC-layer latency for high-throughput devices, while HARQ adapts retransmissions to channel conditions.HARQ can retransmit incremental redundancy for noisy channels or restart transmission under strong interference.

2) MAC-aware Routing Algorithms:

Latency-aware routing must account for interactions between medium access, routing, virtualization, and multiple network domains rather than optimizing each component independently.

  • 2) MAC-aware Routing Algorithms:: MAC parameters influence routing-path performance, while routing-dependent traffic distribution feeds back into MAC behavior.The Q-metric measures contention without queue measurements and can adapt routing for low-traffic, latency-sensitive networks.
  • 2) MAC-aware Routing Algorithms:: Delay Guaranteed Routing and MAC jointly fixes routes and TDMA slots from node positions, eliminating retransmissions under its design.The approach integrates routing and medium access rather than using a separate routing mechanism.
  • 3) NFV and SDN:: NFV and SDN do not reduce latency alone, but provide configurable infrastructure for algorithms that target low-latency operation.SDN can tune NFV blocks and virtual network slices for different latency priorities.
  • V. INTER-NETWORK TECHNIQUES AND TECHNOLOGIES: Multi-domain latency control is harder because intermediate-network delays are often unknown and available only as probability distributions.The paper reviews latency-aware routing and emphasizes content placement and edge caching for cross-domain paths.
  • A. Inter-domain Routing: Path diversity and latency-aware metrics can improve multi-domain load balancing and reduce end-to-end latency.Extended BGP propagates multiple QoS metrics, advertises multiple paths, and selects multiple routes.

B. Edge Caching and Content Placement

Routing optimization cannot overcome physical distance when content is remote, so edge caching and fog computing move content or computation closer to users while introducing placement and resource-management challenges.

  • B. Edge Caching and Content Placement: A ten-domain path adding at least 10 ms per domain has a 200 ms round-trip lower bound, beyond many low-latency targets.Caching content at local edge routers can expand the feasible latency region when routing alone cannot meet the target.
  • B. Edge Caching and Content Placement: Edge caching reduces inter-domain routing delay, improves local service rate, and offloads the main server for latency-sensitive services.Local users contend for cached content rather than waiting for backhaul and inter-domain routing.
  • B. Edge Caching and Content Placement: Effective caching requires demand prediction, content placement, cache dimensioning, and low-complexity distributed policies.Caches are physically distributed while requests arise throughout the network.
  • C. Fog Computing: Fog computing extends cloud resources toward the edge by allowing connected devices to provide computing, storage, and networking functions.Its principal challenges include balancing fog–cloud load and integrating heterogeneous devices.
  • VI. STANDARDS FOR LOW-LATENCY AND ULTRA-RELIABLE COMMUNICATIONS: Current and emerging cellular, industrial, and WLAN standards provide the communication foundations for the low-latency use cases reviewed.The paper analyzes standards for both intra-network and inter-network support.

A. The 3GPP New Radio and 5G Initiative

Low-latency networking combines 5G, industrial, WLAN, unlicensed-spectrum, and time-sensitive Ethernet technologies to reduce transmission and access delays while supporting demanding applications. These approaches target high rates, bounded latency, reliability, and scalable connectivity through coordinated PHY, MAC, and network designs.

  • A. The 3GPP New Radio and 5G Initiative: 5G targets 1 ms latency alongside 10 Gbps peak speed and 100 billion connections, with URLLC addressing mission-critical applications.URLLC is intended for applications including factory automation, smart grids, and intelligent transportation.
  • B. Industrial Communication Networks: Wireless HP targets multi-Gbps aggregate rates, reliability from 10^-6 to 10^-9, and packet transmission below 1 µs.Its PHY design reduces preamble length and optimizes OFDM parameters for short packets.
  • B. Industrial Communication Networks: IEEE 802.15.4e bounds medium access through time-slotted contention-free and contention-based modes, flexible slots, and channel hopping.These mechanisms support periodic and aperiodic traffic while reducing retransmission latency.
  • A. The 3GPP New Radio and 5G Initiative: Unlicensed-spectrum approaches use traffic offloading or fully unlicensed operation to boost data rate and reduce transmission time.License Assisted Access and LTE in Unlicensed Spectrum retain control signaling in licensed bands, whereas MulteFire does not.
  • A. The 3GPP New Radio and 5G Initiative: Time-Sensitive Networking targets deterministic Ethernet latency of 4 µs or less per hop for short messages at 1 Gbps.It schedules time-sensitive data for real-time streaming and industrial or automotive control.

VII. FURTHER DISCUSSIONS

Further progress requires jointly designing short-packet transmission, PHY and MAC parameters, routing, traffic offloading, and mmWave mechanisms. The paper emphasizes sharp trade-offs among latency, reliability, throughput, energy, interference, and mobility-related constraints.

  • VII. FURTHER DISCUSSIONS: Short packets require joint payload and metadata coding and reduced MAC control overhead because channel establishment can exceed transmission time.The paper identifies these issues as open research problems for improving spectral efficiency and latency.
  • VII. FURTHER DISCUSSIONS: Latency can grow sharply beyond performance boundaries, so trade-offs among latency, throughput, reliability, and energy consumption must be characterized.The relevant boundaries depend on the combination of techniques and parameter settings.
  • C. mmWave: mmWave beamforming enables directional high-rate transmission but can itself add delay, while short coherence times complicate beamforming and tracking.Interference may require multilevel HARQ, and prolonged interference makes retransmission less useful.
  • C. mmWave: Blockage may require switching to another mmWave base station or microwave, leaving reassociation and band selection as open problems.Waiting for a blocked main beam is unsuitable when vehicle or pedestrian movement outlasts the PHY frame duration.
  • VII. FURTHER DISCUSSIONS: Contention-based MAC design must combine PHY parameters with MAC parameters because poor SINR can require retransmissions in addition to collision recovery.The paper calls for more effective cross-layer designs for low-latency communication.
  • VII. FURTHER DISCUSSIONS: Routing research must balance directional-link setup time against routing-delay improvements under changing, incompletely known topologies.Directional communication may activate multiple links without interference, but its establishment cost remains important.
  • VII. FURTHER DISCUSSIONS: 5G’s unified infrastructure must serve high-rate media, massive machine-type communication, and latency-sensitive applications despite competing requirements and scale.Traffic offloading can relieve mobile-network load, but latency-sensitive transmission may still lack guarantees across very large connection volumes.

F. Age-of-Information

The paper connects low-latency networking with information freshness, device capabilities, proximity communication, and security overhead. It concludes that end-to-end latency must be decomposed and jointly optimized across layers while respecting reliability, throughput, and other constraints.

  • F. Age-of-Information: Age of Information measures freshness at the destination and depends on both packet delay and packet inter-delivery time.Low delay alone may not ensure good AoI; low queueing delay can coexist with high AoI when arrivals are sparse.
  • G. Hardware and Smart Devices: Larger device caches can proactively store popular content, potentially reducing delay when users request cached items.The paper also identifies smarter processing and caching as device capabilities relevant to low-latency communication.
  • G. Hardware and Smart Devices: D2D communication can use proximity, lower power, and fewer hops, but control and channel-estimation overhead must be weighed against saved delay.The hardware and protocol support required for D2D remains an open research direction.
  • G. Hardware and Smart Devices: Security and privacy mechanisms contribute substantially to communication and processing delay, while cryptography adds heavy coordination overhead.The paper identifies alternative low-latency privacy methods as requiring substantial investigation.
  • VIII. CONCLUSION: The paper’s conclusion recommends decomposing end-to-end delay across physical-to-transport layers and jointly optimizing techniques under reliability and throughput constraints.It identifies MU-MIMO, mmWave, and full-duplex as enabling technologies that also introduce challenges such as beamforming delay.
Loading 1808.02079v1…