Boosting 5G-New Radio reliability: The power of cognitive packet duplication

Ultra-Reliable Low-Latency Communication and its use cases

5G can support a diverse range of use cases. One fundamental class of use cases that can be supported in 5G is ultra-Reliable Low-Latency Communication (uRLLC). Ultra-high network reliability (>99.999%) and extremely low latency (~1ms) in packet transmission address use cases that demand near-instantaneous data transmission and minimal error rates. Among those use cases, we find applications in healthcare (for example, Augmented Reality-assisted or remote surgery), intelligent transportation, drone control, smart factory or industrial automation, smart electricity grid, and port automation. Given the massive market prospects, prioritizing reliability becomes crucial to meet the requirements of 5G uRLLC use cases.

The practical implementation of uRLLC is often limited by network-related challenges such as congestion, radio environment, power considerations, mobility requirements, or the limitations of existing standards and protocols. It is, therefore, imperative to proactively manage uRLLC use cases, and Artificial Intelligence/Machine Learning (AI/ML) techniques have been applied to address the challenges faced by the technology. Cognitive Packet Data Convergence Protocol (PDCP) packet duplication, which we discuss in this post, is one such AI/ML technique.

To achieve system-wide integrated support for uRLLC, 3GPP (The 3rd Generation Partnership Project) has rolled out, since release 15, the PDCP packet duplication standard. PDCP packet duplication uses two independent radio paths to transmit the same data, which, by definition, increases redundancy. Redundant links increase reliability, while low latency is achieved by eliminating the need for retransmitting packets.

Figure 1: PDCP packet duplication in 5G NSA.

Figure 2: PDCP packet duplication in 5G SA.

Figures 1 and 2 show different approaches for PDCP packet duplication in 5G NSA (NonStandalone) and SA (Standalone) modes. In 5G NSA, a split bearer—a bearer is a dedicated communication channel—can be employed to transmit the same packet using primary (master) and secondary node links. In contrast, in 5G SA mode, Carrier Aggregation (CA) can be used to transmit the same information across different frequency carriers.

In this post, we discuss how the use of AI/ML techniques and approaches can be applied for the selective activation of PDCP packet duplication over Always-ON activation mode. Our focus is on specific ML models that can help predict and determine cases where packet duplication is beneficial to the network.

Existing PDCP packet duplication: the challenges

The purpose of PDCP Packet Duplication (PD) is to enhance redundancy, although this capability might not consistently provide advantages over the entire operational lifespan of a bearer. For ease of deployment for URLLC use cases, numerous implementations suggest enabling packet duplication by default. Other existing implementations might instead rely solely on Reference Signal Received Power (RSRP) to determine whether packet duplication can be enabled, without factoring in additional performance criteria.

Under favorable channel conditions (for example, strong signal, low noise, and interference), PD can lead to an inefficient utilization of air interface resources. In such advantageous conditions, a single link transmission might meet the required levels of latency and reliability. As such, activating PD without a careful assessment of the duplication gain could result in needless overhead—where overhead denotes the occupation of additional air interface resources—which ultimately diminishes network capacity.

In the context of 5G NSA deployment, PD is implemented by using resources from the primary anchor LTE node and secondary 5G node resources (Fig 1). On the other hand, in 5G SA configurations, PD can be achieved through carrier aggregation. With PD CA, identical user data is transmitted across multiple carriers simultaneously (Fig 2). PD allocates resources from a secondary carrier (CA component carrier) or secondary node (5G NSA gNodeB) specifically for duplicating transmissions and handling retransmissions. This allocation can have implications for the original scheduling of uRLLC in the component carrier or secondary node.

Emerging hybrid use cases, including those focusing on high throughput like Enhanced Mobile Broadband (eMBB)—and particularly ultra-High Speed and Low Latency Communication (uHSLLC) applications combining eMBB with URLLC, demand a combination of high throughput (meaning a high amount of data successfully transmitted within a specified time frame), low latency, and robust reliability. Therefore, the efficient and adaptive utilization of PDCP duplication becomes crucial to ensure the fulfillment of use case requirements on capacity, latency, and reliability.

ML-based PDCP packet duplication (PD) instance identification

To effectively manage the activation of dynamic packet duplication, ensuring a positive customer experience and meeting the requirements of uRLLC, it is important to monitor critical performance factors such as network reliability, traffic patterns, packet errors, and performance metrics: for example, latency, throughput, and Channel Quality Indicator (CQI) must be proactively monitored.

Figure 3 shows, at a high level, a proposal for a cognitive packet duplication approach. The entire solution includes the following four sequential components:

1. Input parameters
2. Cognitive layer
3. ML Predictions (or inference)
4. Resulting Actions

Figure 3: The high-level workflow for cognitive packet duplication decision.

Here below, we provide more details on the four components.

The main input parameters for this cognitive solution are:

1. User Equipment (UE) mobility state: The solution starts with classifying UEs into static or mobile states. Static UEs remain in similar radio channel conditions for long and do not face high variations. UEs, which are moving might encounter varying radio conditions, therefore such UEs will be prone to degraded reliability and latency.
2. UE capabilities: the capabilities that are supported by UEs. For example, whether UE is 5G SA or NSA capable.
3. 5G deployment mode: The identification of the optimal packet duplication action is determined by considering both the capabilities of UE and the network topology in SA/NSA scenarios.
4. Requested service: The requested service and associated QoS flows for Data Radio Bearers (DRB) level PD activation.
5. Performance attributes: To provide network performance data and predictive network statistics, a range of metrics are used as inputs. These metrics include factors and indicators like latency, throughput, buffer size, radio link failures, timing advance, cell utilization, CQI, and Uplink (UL) radio conditions.

The cognitive layer comprises multiple machine learning models or optimizers, which can be selected and used based on the available data type and modeling objective. These might include:

1. Regression models, such as those applied for predicting performance metrics or the location for PD. Such models might take as input the radio data bearer, UE velocity, CQI, handover count to neighbor cells, SINR, and other network parameters. Among such models, we might also use Recurrent (or Deep) Neural Network models that capture temporal dependencies of parameters to determine the location and predicted metric output.
2. Federated or Distributed Learning models are models trained on local datasets of network parameters. Subsequently, model parameters or coefficients are aggregated at a central node. Local models might undergo updates at specified time intervals.
3. Reinforcement learning models, in which a reward function is used to determine actions that control metrics in a state-action space. The state might include RSRP, traffic, and gNB location, among other parameters, while the action might be turning on or off PD [3,4,5].

The ML Predictions component serves the output from the cognitive layer, which will help take the appropriate PD decision. The deployed ML models would help predict UE locations where use case specific SLAs would not be fulfilled.
The cognitive layer might also predict intra/inter-frequency handover (HO) probability at the predicted location, RSRP, packet error loss, BLER (block error rate), and Xn interface latency for anchor eNodeB and gNodeB. Also, carrier aggregation evaluation regarding packet multiplexing performance vs packet duplication would be done using input parameters to identify the PD gain or acceptable tradeoffs in network performance.

Actions, which are based on the ML predictions, will assess whether enabling PDCP packet duplication is necessary and whether standard network operations, such as intra or inter-frequency handovers (HOs), can address the PD requirement.
If the handover (HO) process alone does not meet the required level of reliability, the suitable method for activating Packet Duplication (PD) and the number of instances for activation is based on the available action space. The actions to be taken are determined by a balanced approach of redundancy, reliability, and resource utilization. The objective remains to maximize reliability and spectral efficiency by selectively activating the PD only when it is needed.

References [3,4,5] show the potential benefit of Machine Learning based packet duplication, where a significant reduction (up to 81%) in packet duplication is observed while maintaining the latency requirements.

The cognitive layer might reside in the base station (LTE eNB/5G gNB), with limited capabilities and footprint, or could be part of a centralized platform, such as the Ericsson Intelligent Automation Platform (EIAP), which is a type of Service Management and Orchestration Platform (SMO). 3GPP has introduced a Network Data Analytics Function (NwDAF) that can also perform various analytics based on different data sources. Since the target use case is the uRLLC, which has very strict latency requirements, the proximity of the solution to users (base stations) will help minimize the latency of decisions, and it, therefore, appears to be the preferred location for the cognitive layer.

Looking Forward

Given the complexity of 5G systems, AI/ML techniques can be of great help when optimal decisions are needed at the sub-millisecond level. The cognitive PDCP packet duplication stands out as an ML-based paradigmatic solution that considers network factors and indicators, such as mobility and radio environment, for the deployment of PD. This solution can help achieve a reduction of instances of packet duplication within the 5G system while maintaining the required level of latency and reliability.

At Ericsson, we look forward to collaborating with communication service providers (CSPs) and technology partners that are interested in deploying uRLLC services to build the autonomous networks of the future.

Summary

A cognitive layer-based decision to enable packet duplication helps minimize redundant air interface resource usage. Cognitive layer-based PD enablement assists in proactively identifying spatial location and time instances at which latency or packet loss might negatively affect UE performance. Furthermore, the cognitive layer helps evaluate various potential network actions before recommending PD, thus pinpointing the optimal action for maintaining high network performance. Enabling packet duplication dynamically on a per-need basis makes the network achieve a similar level of latency performance to the always ON mode, with the crucial advantage that reducing packet duplication significantly resulting in better resource utilization.

Learn more

5G URLLC - Achieving industrial automation

Realizing Time-Critical Communication with 5G

Sustainable networks: Saving energy while maintaining customer experience

Explore 5G.

Cognitive Reasoning for 5G Network Lifecycle Management