Mitigation of the noisy neighbor problem in VRAN deployments
The described method optimizes vRAN deployments by using machine learning to configure distributed unit and central unit placements and functional splits, effectively mitigating the noisy neighbor problem, reducing computing overhead and latency, and enabling scalable edge deployments.
Patent Information
- Application Number
- PCT/IB2023/057107
- Authority / Receiving Office
- WO · WO
- Patent Type
- Applications
- Current Assignee / Owner
- Priority Date
- 2023-05-04
- Filing Date
- 2023-07-11
- Publication Date
- 2024-11-07
AI Technical Summary
The noisy neighbor problem in vRAN deployments leads to increased computing overhead, resource contention, energy consumption, and latency due to non-ideal isolation between virtualized base stations sharing computing resources at the edge, which existing methods like cache allocation and resource isolation struggle to fully mitigate, especially in large-scale deployments.
A computer-implemented method and system that jointly determines the distributed unit computing configuration, central unit placement, and functional split configuration using machine learning to minimize the noisy neighbor effect by optimizing the placement and workload distribution between central and distributed units, allowing for flexible configuration of RAN network functions.
This approach reduces computing overhead and latency while maintaining quality of service, enabling scalable vRAN deployments even in resource-constrained edge environments, conserving computational resources, and minimizing capital expenditures.
Smart Images

Figure IB2023057107_07112024_PF_FP_ABST
Abstract
Description
MITIGATION OF THE NOISY NEIGHBOR PROBLEM IN VRAN DEPLOYMENTSCROSS-REFERENCE TO RELATED APPLICATION
[0001] Priority is claimed to U.S. Provisional Patent Application No. 63 / 500,000, filed on May 4, 2023, the entire disclosure of which is hereby incorporated by reference herein.FIELD
[0002] The present disclosure relates to a method, system, controller and computer-readable medium for virtualized radio area networks (vRANs), in particular, for mitigating the noisy neighbor problem in vRAN deployments.BACKGROUND
[0003] The virtualization of the radio access networks (RANs) is the next step toward a fully virtualized mobile network. RAN virtualization softwarizes the RAN functionalities into a common computing platform reducing costs by running them on generic hardware and enabling multiple vRANs to share a common pool of resources. vRANs provide increased flexibility by making it simpler to rollout updates and perform control and management operations.
[0004] The Open RAN Alliance (O-RAN) stands among the initiatives aiming to define the specifications of the next generation of vRANs. The O-RAN architecture distributes the radio functions into three main building blocks: (i) the radio unit (RU), (ii) the distributed unit (DU), and the (iii) central unit (CU). See O-RAN Alliance, “O-RAN-WG1 -O-RAN Architecture Description - v04.00.00,” Tech. Spec. (March 2021), the entirety of which is hereby incorporated by reference herein.SUMMARY
[0005] An aspect of the present disclosure provides a computer-implemented method for configuring virtual base stations in a network having an edge server and a platform server. The edge server and the platform server host the virtual base stations. Each respective virtual base station of the virtual base stations has a respective central unit and a respective distributed unit. The method includes: for each of the virtual base stations, jointly determining a respective distributed unit computing and functional split configuration and a respective central unit placement configuration to minimize a noisy neighbor problem based on a set of metrics collected from the virtual base stations.BRIEF DESCRIPTION OF THE DRAWINGS
[0006] Embodiments of the present disclosure will be described in even greater detail below based on the exemplary figures. Aspects of the present disclosure are not limited to the exemplary embodiments. All features described and / or illustrated herein can be used alone or combined in different combinations in embodiments of the present disclosure. The features andadvantages of various embodiments of the present disclosure will become apparent by reading the following detailed description with reference to the attached drawings, which illustrate the following:
[0007] FIG. 1 illustrates the O-RAN architecture;
[0008] FIG. 2 illustrates different O-RAN deployment options;
[0009] FIG. 3 illustrates different possible functional split configurations for vRAN deployments;
[0010] FIG. 4 illustrates a method and system for vRAN configuration and deployment according to an embodiment of the present disclosure; and
[0011] FIG. 5 illustrates the different functional blocks of a controller according to an embodiment of the present disclosure.DETAILED DESCRIPTION
[0012] Embodiments of the present disclosure provide a method, system, controller and computer-readable medium that mitigate the computing overhead caused by multiple virtualized base stations sharing computing resources at the edge in a multi-tenant vRAN deployment, a problem known as the noisy neighbor problem. In particular, embodiments of the present disclosure provide a mechanism to minimize the overall computing overhead experienced at the edge due to the noisy neighbor problem by deciding the placement, the computing resources, and the functional split (if supported) of the RAN network functions.
[0013] FIG. 1 illustrates the O-RAN architecture 100 for a 5G core network 140 including the radio unit (RU) 110, distributed unit (DU) 120, and central unit (CU) 130. The radio unit (RU) 110 contains the radio hardware (RF) 112 and handles the lower physical (PHY) layer 114 functionalities, which convert radio signals sent to (or received from) the antenna 102 into digital radio samples that are then sent to (or received from) the distributed unit (DU) 120. The distributed unit (DU) processes the digital radio samples from the radio unit (RU) 110 and performs part of the RAN functions, such as the upper PHY layer 122, medium access control (MAC) layer 124, and radio link control (RLC) layer 126. Finally, the central unit (CU) 130 handles the highest layer functionalities, such as the radio resource control (RRC) layer 132 and packet data convergence protocol (PDCP) layer 134.
[0014] O-RAN defines different deployment options for each of the logical functions introduced above. See, e.g., O-RAN Alliance, “Cloud Architecture and Deployment Scenarios v02.01 (ORAN. WG6.CAD-v02.01),” Tech. Rep. (July 2020), the entirety of which is hereby incorporated by reference hereError! Reference source not found.. Typically, the radio units (RUs) are deployed into a cloud cell site, the distributed units (DU) are deployed into an edge computing cloud site and the central units (CUs) are deployed into a cloud platform. However,there are other options. For example, the distributed unit (DU) and radio unit (RU) can be either co-located at the same cloud cell site or separated by placing the radio unit (RU) into a cell site and the distributed unit (DU) into the edge cloud platform. O-RAN defines different placement options including proprietary cell and edge sites, which provide limited functionalities to the network operators.
[0015] FIG. 2 shows five different deployment scenarios 240 for the radio unit (RU) 210, distributed unit (DU) 220, and central unit (CU) 230 of an O-RAN architecture for a 5G core network.
[0016] In scenario A 241, one edge cloud 247 centralizes all virtual central unit (CU) 230 and distributed unit (DU) 220 functions. The radio unit (RU) 210 is deployed a proprietary cell site 246.
[0017] In scenario B 242, the central unit (CU) 230 functions are located in a cloud platform 248, which is different from edge cloud platform 247 where the distributed unit (DU) 220 is deployed. The radio unit (RU) 210 is deployed a proprietary cell site 246.
[0018] In scenario C 243, the distributed unit (DU) 220 and the radio unit (RU) 210 functions are located in a proprietary edge site 249 and proprietary cell site 246, respectively. The central unit (CU) 230 functions are located in a cloud platform 248.
[0019] In scenario D 244, both the distributed unit (DU) 220 and the radio unit (RU) 210 functions are located in a cloud cell site 250. The central unit (CU) 230 functions are in a separate cloud platform 248. Some edge hardware and low level processing of the radio unit (RU) 210 may be located at a proprietary cell site, depending on the instances of a deployment under this scenario.
[0020] In scenario E 245, the central unit (CU) 230, distributed unit (DU) 220, and radio unit (RU) 230 functions are all separated into different cloud locations: cloud platform 248, cloud edge site 247, and cloud cell site 250, respectively.
[0021] The present disclosure uses terms consistent with how a person of ordinary skill in the art would understand those terms to be used in the context of O-RAN and vRAN. For the sake of clarity, however, several definitions are given below:
[0022] Radio unit (RU): A radio unit (RU) refers to the component in O-RAN system that is responsible for transmitting and receiving radio signals, and hosting the low-PHY layer functions based on the lower layer functional split. It includes radio frequency hardware (RF) components and baseband processing capabilities. The RU converts digital signals into radio waves for wireless transmission and vice versa. In the O-RAN architecture, the RU is located at the edge of the network, closer to the antennas or includes the antennas.
[0023] Distributed unit (DU): The distributed unit (DU) is a network element in O-RAN that is responsible for processing the high-PHY, MAC and RLC layers from the IQ samples received from one or more radio units (RUs). It is located (functionally) closer to the RUs. The DU communicates with the RUs through the front haul links to exchange IQ samples and the central unit (CU) to coordinate and manage the different traffic flows of the users. The DU can be seen as an intermediate layer between the radio units (RUs) and the central unit (CU).
[0024] Central unit (CU): The central unit (CU) is a network element in O-RAN that processes the higher layers of the radio stack as the RRC and PDCP layers . It is responsible for managing the life cycle of the connection and the QoS of the different traffic flows. The central unit (CU) interacts with the distributed units (DU) to manage user connections, and ensure efficient operation of the network.
[0025] Lower physical (PHY) layer: The lower physical (PHY) layer refers to the lowermost layer in the protocol stack of a virtual radio access network (vRAN) system. It is responsible for transmitting and receiving raw physical signals over the wireless medium. The lower PHY layer deals with tasks such as modulation, coding, and channel multiplexing, ensuring reliable transmission of data between the radio units (RUs) and distributed units (DUs).
[0026] Upper physical (PHY) layer: The upper physical (PHY) layer is the higher portion of the PHY layer in the protocol stack of a vRAN system. It builds upon the lower PHY layer and provides functions such as channel quality measurement, channel selection, and higher-level signaling related to physical layer control. The upper PHY layer helps optimize the transmission parameters for better performance and efficiency.
[0027] Medium access control (MAC) layer: The medium access control (MAC) layer resides above the PHY layer in the vRAN protocol stack, and is responsible for managing access to the shared radio resources. The MAC layer handles tasks such as scheduling, prioritization, and coordination of data transmissions between multiple users or devices connected to the vRAN network.
[0028] Radio link control (RLC) layer: The radio link control (RLC) layer is a layer above the MAC layer in the vRAN protocol stack. It provides reliable and error-free transmission of data across the radio link. The RLC layer handles functions like segmentation and reassembly of data packets, flow control, error detection, and retransmissions to ensure data integrity and quality of service.
[0029] Radio resource control (RRC) layer: The radio resource control (RRC) layer is a higher-level layer in the vRAN protocol stack responsible for controlling and managing radio resources. It handles tasks such as connection establishment, configuration, and release, as wellas mobility management. The RRC layer facilitates efficient utilization of radio resources and ensures seamless handovers between cells or base stations.
[0030] Packet data convergence protocol (PDCP) layer: The packet data convergence protocol (PDCP) layer is located above the RLC layer in the vRAN protocol stack. It provides header compression, encryption, and integrity protection for packet data transmission. The PDCP layer is responsible for optimizing the transmission efficiency and security of internet protocol (IP) packets over the radio interface, enabling efficient utilization of the wireless network resources.
[0031] 5G core: The 5G core refers to the central component of a fifth-generation (5G) network architecture that handles the core network functions. It is responsible for providing various network services, managing user connections, and facilitating the communication between different network entities. The 5G core interacts with the vRAN protocol stack to enable end-to-end communication and support advanced 5G functionalities. Some of the functionalities provided by the 5G core include: network functions virtualization infrastructure (NFVI); network slice management and orchestration (NSMO); session management function (SMF); user plane function (UPF); policy control function (PCF); and authentication and authorization function (AAF).
[0032] Proprietary cell site: A proprietary cell site refers to a cellular network infrastructure deployment that is owned and operated by a specific service provider or organization. It typically consists of base stations, antennas, and related equipment used to provide wireless coverage and connectivity within a specific geographical area. Proprietary cell sites are usually built and maintained by individual network operators to support their own cellular networks.
[0033] Proprietary Edge Site: A proprietary edge site refers to a localized computing infrastructure that is privately owned and operated by a specific organization. It is typically deployed at the network edge or close to the end-users or devices (e.g., base station at the proprietary cell site) it serves. A proprietary edge site may consist of computing resources, networking equipment, and storage devices that are dedicated to supporting specific applications or services. It offers the advantages of localized processing, reduced latency, and enhanced control for organizations that require specific performance or security requirements.
[0034] Cloud Platform: A cloud platform refers to a comprehensive and scalable infrastructure that provides a wide range of remote computing services and resources over at network (e.g., over the internet). It typically includes virtualized servers, storage, networking capabilities, and software services that can be accessed and utilized on-demand. Cloud platforms allow users to deploy, manage, and scale applications and services without the need for local-dedicated physical infrastructure. They provide flexibility, scalability, and cost-efficiency by enabling resource pooling and self-service provisioning.
[0035] Cloud Cell Site: A cloud cell site refers to a virtualized and cloud-based deployment model for cellular network infrastructure. Instead of physical hardware deployed at each cell site, the functions of a traditional cell site are virtualized and hosted on cloud platforms. Cloud cell sites leverage the benefits of cloud computing, such as resource elasticity, centralized management, and efficient resource allocation. By virtualizing the cell site functions, cloud cell sites can improve scalability, flexibility, and cost-effectiveness in managing cellular network deployments.
[0036] Edge cloud: An edge cloud is a virtualized and cloud-based deployment model for network edge functionality. Thus, it is computing infrastructure (generally distributed over a network) that brings remote computing resources at or close to the network edge (from a functional standpoint). It involves deploying remote computational services and resources at or near the edge of the network, e.g., closer to the device deployment (e.g., cell site hardware).
[0037] Although vRANs offer a number of benefits, there is a high price to be paid in terms of computing overhead due to non-ideal isolation between virtual distributed units (DUs) deployed at the edge cloud. This can result in frequent resource contentions and service degradation, known as the noisy neighbor problem. The noisy neighbor problem also negatively affects energy consumption and latency, as resource contention increases the number of computing cycles used to access the resources.
[0038] To reduce the noisy neighbor problem, some common methods used by existing technology include assigning virtual base station (vBS) processes to different computing cores or using tools like INTEL cache allocation technology (CAT) to allocate different lines of cache memory for different computing cores. However, these methods may not be feasible for large- scale deployments at the network edge where computing resources are limited, and there may be more vBSs than computing cores or cache lines. As a result, operators may have to accept some level of the noisy neighbor effect. While most operators have relied on isolating different computing resources to address this issue, there are other strategies that can also be effective in mitigating the computing overhead.
[0039] On the one hand, the placement of the different components can have a significant impact on mitigating the noisy neighbor problem since placing the central unit (CU) in a different location than the distributed unit (DU) reduces the total computing demand in the edge cloud platform. However, there is a trade-off since co-locating the distributed unit (DU) and the central unit (CU) reduces the total latency experienced by the user as the data passed between the distributed unit (DU) and the central unit (CU) does not experience a significant latency dueto the transport network. On the other hand, the functional split between the central unit (CU) and distributed unit (DU) can significantly affect the computing consumption in a vRAN system. If the split is configured so that the distributed unit (DU) is responsible for more processing tasks, it can lead to higher computing consumption. This is because the distributed unit (DU) will require more processing power to handle the complex signal processing tasks and the upper layer computations. Nevertheless, if the split is configured so that the central unit (CU) is responsible for more processing tasks, it can lead to lower computing consumption in the distributed unit (DU). The configuring of the functional split, however, only makes sense when the distributed unit (DU) and the central unit (CU) are located in different hosts. Changing the split whenever they are co-located will not change the computing demand, as all functions will still run on the same computing platform.
[0040] Although O-RAN‘s functional split between the central unit (CU), distributed unit (DU) and radio unit (RU) is fixed, other vRAN implementations (including proprietary vRAN implementations) may allow for additional flexibility in deciding the placing of some functional layers of the vRAN stack.
[0041] FIG. 3 illustrates, for example, three different possible functional split configurations for vRAN implementations.
[0042] In Configuration 1 300, the split configuration separates the upper PHY layer 322 functions processing from the higher layer processing. In particular, the central unit (CU) 330a is responsible for higher-level functions such as scheduling and radio resource management (e.g., RRC layer 332, PDPC layer 334, RLC layer 326, and MAC layer 324 functionalities), while the distributed unit (DU) 320a is responsible for the upper PHY layer 322 functions. The radio unit (RU) 310 is responsible for the lower PHY layer 314 functions and includes the radio hardware (RF) 312.
[0043] In Configuration 2 301, the split configuration is similar to previous one, but it also splits the MAC layer 324 processing between the central unit (CU) 330b and the distributed unit (DU) 320b. The central unit (CU) 330b is responsible for higher-level functions such as radio resource management and scheduling (e.g., RRC layer 332, PDPC layer 334, and RLC layer 326 functionalities), while the distributed unit (DU) 320b is responsible for low-level processing such as data encoding and decoding, as well as MAC processing (e.g., upper PHY layer 322 and MAC layer 324 functionalities).
[0044] In Configuration 3 303, the upper PHY layer 322 functions, MAC layer 324, and the RLC layer 326 processing are split between the central unit (CU) 330c and the distributed unit (DU) 320c. The central unit (CU) 330c is responsible for higher-level functions such as thePDCP layer 334 and the RRC 332, and the distributed unit (DU) 320c processes the RLC layer 326, the MAC layer 324, and the upper PHY layer 322 functions.
[0045] Proprietary implementations can support more options than the ones described above in FIG. 2. On the other hand, FIG. 2 shows the only currently specified deployment options for O-RAN compliant components. Changing the functional split configuration between the O-RAN components essentially changes the workloads of the distributed unit (DU) and the central unit (CU) as the memory and computing requirements are different.
[0046] Embodiments implemented according to aspects of the present disclosure overcome the limitations of existing technology to mitigate the noisy neighbor problem occurring at the edge cloud computing platform in a vRAN deployment. Aspects of the present disclosure provide a machine learning mechanism to determine an objectively optimized central unit (CU) placement, computing configuration, and the distributed unit (DU) functional split configuration with the central unit (CU) (if supported). On the one hand, co-locating fewer distributed unit (DU) and central unit (CU) instances together decreases the aggregated host computing consumption because fewer instances share the same computing resources. This reduces contention when accessing the shared computing resources. However, processing some radio functions in the cloud potentially increases the total latency of a vRAN deployment, which is critical to provide synchronization between RAN network functions. On the other hand, offloading more functions to a central unit (CU) placed into a different location configuring the functional split reduces the total computing demand of the distributed units (DUs) deployed in the edge which could also help mitigate the noisy neighbor problem.
[0047] Thus, given a scenario where there are multiple distributed units (DUs) deployed in an edge computing server experiencing increased computation times due to the noisy neighbor problem, embodiments of the present disclosure enable configuring the distributed unit (DU) computing configuration, the central unit (CU) placement (either collocating them with the corresponding distributed units (DUs) or in the cloud) and the functional split between the distributed unit (DU) and the central unit (CU) (if supported) such that the noisy neighbor effect is objectively minimized without significantly increasing the overall computational burden.
[0048] In an embodiment, the present disclosure provides a computer-implemented system to decide the central unit (CU) placement jointly with the distributed unit (DU) computing configuration in the edge cloud and their functional split configuration (if supported) in an O- RAN deployment so that the total computing consumption and the effect of the noisy neighbor’s problem is reduced or minimized. Embodiments of the present disclosure provide improvements over existing technology that tries to reduce the computing overhead caused by the noisy neighbor problem by considering the placement of the distributed unit (DU) and central unit(CU) and their functional split to further or more significantly reduce the computation overhead. Existing technology only considers resource allocation of the computing resources, which creates a problem for large-scale deployments where the operators might be forced to share resources in the edge-computing host.
[0049] In a first aspect of the present disclosure, a controller is provided that optimizes the placement of the central units (CUs) and in an 0-RAN deployment, as well as their functional split configuration (if supported), j ointly with the computing configuration of the distributed units (DUs) deployed in the edge-computing platform, to thereby mitigate the noisy neighbor effect in the edge cloud host of a vRAN deployment.
[0050] A second aspect off the present disclosure provides a method for configuring a vRAN deployment which includes the following operations. The previous listed metrics for each user server by each distributed unit (DU) deployed in the edge cloud platform are fetched and saved to a database that is accessible by the controller. If it is not possible to access user data, the metrics agent collects information in the form of aggregated traffic and a number of users a distributed unit (DU) is serving. The control interfaces are then accessed to update the placement and computing configuration of the instances deployed. If supported, the interfaces that allow changing the functional split configuration are also accessed. The optimal computing, placement and functional configuration of the different central unit (CU) and distributed unit (DU) components for each vBS deployed are computed to minimize the noisy neighbor effect as the workloads of the distributed unit (DU) and the central unit (CU) change when the functional split is configured. The computed configuration can be enforced through control application programming interfaces (APIs). In some implementations of the present aspect, a feedback signal is used to learn the latent relationship between the context input and the configuration of the vBSs deployed.
[0051] In the context of the present disclosure, minimizing the noisy neighborhood effect means determining which configuration of the central unit (CU) and distributed unit (DU) from among a set of possible configurations results achieving the lowest predicted noisy neighbor effect as compared to the other possible configurations. The possible configurations may be restricted based on the computing resources available, maintaining a minimum quality of service (QoS), and / or maintaining a predefined maximum latency. The possible configurations may be further limited based on having to jointly determine two or more configurations, including distributed unit computing configurations, central unit placement configurations, and / or functional split configurations. The possible configurations may be still further limited based on having to jointly determine multiple configurations for multiple instances of pairs of central units (CUs) and distributed units (DUs).
[0052] In a third aspect of the present disclosure, there is provided a computer-implemented method for configuring virtual base stations in a network having an edge server and a platform server. The edge server and the platform server host the virtual base stations. Each respective virtual base station of the virtual base stations has a respective central unit and a respective distributed unit. The method includes: for each of the virtual base stations, jointly determining a respective distributed unit computing configuration and a respective central unit placement configuration to minimize a noisy neighbor problem based on a set of metrics collected from the virtual base stations.
[0053] In a fourth aspect, for the method according to the third aspect, for each of the virtual base stations, the joint optimization further jointly optimizes a respective functional split configuration between the respective central unit and the respective distributed unit together with the distributed unit computing configuration and the central unit placement.
[0054] In a fifth aspect, the method of the third or fourth aspect further includes, for each of the virtual base stations, commanding the respective virtual base station to update based on the respective distributed unit computing configuration, and / or the respective central unit placement configuration, and / or the respective functional split configuration.
[0055] In a sixth aspect, the method of any one of third through fifth aspects further includes: generating a plurality of context vectors by generating, for each of the virtual base stations, a respective context vector based on metrics that are associated with the respective virtual base station and which are from the set of metrics collected; and embedding each of the context vectors into a fixed dimension space. The joint optimization includes using a machine learning model to learn latent relationships between the context vectors and actions configurations, and for each of the base stations, the respective actions configurations include actions for at least one of: placing the respective central unit at the edge server or the platform server, configuring computing resources for the respective distributed unit deployed at the edge server, or splitting functionalities between the respective distributed unit and the respective central unit.
[0056] In a seventh aspect, the method of any one of the third through sixth aspects further includes: using the machine learning model, learning a latent relationship between the context vectors and configurations of each of the virtual base stations currently deployed based on a feedback signal. The feedback signal having information on a computing consumption used at the edge server and / or information on function of the virtual base stations.
[0057] In an eighth aspect, the method of any one of the third through seventh aspects further includes: for each of the virtual base stations, retrieving current metrics associated with the respective distribution unit and saving the current metrics as part of the set of metrics. Foreach of the virtual base stations, the current metrics include at least one of a signal -to-noise ratio, traffic demand, or number of active users, the metrics being associated with each user served by the respective distribution unit and / or being aggregated metrics for all of the users served by the respective distribution unit.
[0058] In a ninth aspect, for the method of any one of the third through eighths aspects, using the machine learning model includes: executing a plurality of artificial intelligence (Al) agents to determine the respective actions configurations, each of the Al agents being associated with one of the virtual base stations.
[0059] In a tenth aspect, for the method of any one the third through ninth aspects, the Al agents take as their inputs the context vectors.
[0060] In an eleventh aspect, for the method of any one of the third through tenth aspects, using the machine learning model includes: using a common model that is configured to learn how the Al agents collaborate to select the actions configurations to minimize the noisy neighbor problem, and updating the Al agents based on the common model.
[0061] In a twelfth aspect, for the method of any one of the third through eleventh aspects, the common model uses the feedback signal to learn the latent relationship between the context vectors and configurations of each of the virtual base stations and to provide rewards to the Al agents based on the feedback signal.
[0062] In a thirteenth aspect, for the method of any one of the third through twelfth aspects, using the machine learning model includes: using an attention head to embed different actions and contexts of the different Al agents, the agents being configured to learn how the other Al agents’ actions affects their actions based on the embeddings from the attention head.
[0063] In a fourteenth aspect, for the method of any one of the third through thirteenth aspects, the respective functional split configuration between the respective central unit and the respective distributed unit specifies which of the respective central unit or the respective distributed unit executes one or more of the following functions: medium access control (MAC) layer functions or radio link control (RLC) layer functions.
[0064] According to a fifteenth aspect, there is provided a computer system with one or more hardware processors which, alone or in combination, are configured to provide for execution of a method according to any of the first through fourteenth aspects.
[0065] According to a sixteenth aspect, there is provided a tangible, non-transitory computer-readable medium having instructions thereon which, upon being executed by one or more hardware processors, alone or in combination, provide for execution of a method according to any of the first through fourteen aspects.
[0066] Embodiments implemented according to one or more aspects of the present disclosure provide several improvements over existing technology in the field 5G network configuration, particularly with respect to vRAN deployment and design. What follows are a few of such improvements.
[0067] Aspects of the present disclosure advantageously provide mechanisms for automatically and optimally determining and / or configuring: (i) the central unit (CU) placement; (ii) the distributed unit (DU) and central unit (CU) functional split, and / or (iii) the computing configuration (e.g., in an edge cloud platform) to minimize the noisy neighbor effect. For example, aspects of the present disclosure provide a mechanism that advantageously can automatically, jointly optimize the distributed unit (DU) computing configuration and the central unit (CU) placement, which minimizes the noisy neighbor problem. Also, aspects of the present disclosure implemented in a scenario that allows changing of the functional split between the distributed unit (DU) and the central unit (CU), can provide a mechanism that automatically optimizes the split (alone or together with the automatic configuration / optimization). By changing the split, the workload share among the distributed unit (DU) and the central unit (CU) can be configured because the set of functions that they have to process can be changed.
[0068] Aspects of the present disclosure advantageously provide a controller implementing a machine learning model that uses a set of metrics information (e.g., collected from the distributed units (DUs)) to create a context vector for each user. The context vector can be embedded into a fixed dimension space, and the controller implements the machine learning model to learn the latent relationship between the context vectors and the actions configurations. Thus, an aspect of the present disclosure advantageously provides a controller that teaches different artificial intelligence (Al) agents (which map to different vBSs) how to collaborate in selecting the optimal configuration that minimizes the noisy neighbor effect.
[0069] By automatically optimizing one or more the placement, computing configuration, and / or the functional split of the different vBS deployed in order to reduce the effect of the noisy neighbor problem caused by deploying multiple vBSs in the same edge computing platform, other advantageous effects are also achieved. For example, such optimization leads to less computing overprovisioning in 0-RAN deployments, which translates to less equipment / hardware, which translates into conserving of computational resources, saving computing power, reduced computational burden and a lower capital expenditures — as compared to existing systems, which cannot provide the optimization of the present disclosure. Additionally, aspects of the present disclosure enable scalable vRAN deployments, particularly enabling large scale deployments at a network edge where computational resources are limited, which also maintaining optimal latency in network transportation.
[0070] These and other advantages and features are further elucidated below in connection with exemplary embodiments of the present disclosure.
[0071] For example, FIG. 4 illustrates a method and system 400 for automatic vRAN configuration and deployment according to an embodiment of the present disclosure in a scenario with a mobile network infrastructure 410 having a cloud platform 420, an edge cloud platform 430 , and a cloud cell site 440. There are different vBSs 431 (e g., VBSi, VBS2, VBSs, . . . , VBSN) deployed where the radio units (RUs) are deployed into the cloud cell site 440 and their distributed units (DUs) into the edge cloud platform 430, and the central units (CU) can be either co-located with the distributed units (DU) or deployed into the cloud platform 420. The method and system configure the placement of the central units (CUs), the functional split between the distributed unit (DU) and the central unit (CU), and the computing resources of the distributed units (DUs) in the edge computing host 430. The cloud platform is in network communication with the 5G core 450.
[0072] As used herein, “cloud” refers to one or more computing resources that are hosted and / or managed by one or more servers connected through a network communication infrastructure. A cloud platform can provide access to a wide range of applications, services, and data storage capabilities over the network. The computing resources provided the server typically include one or more processors and storage systems that can be made available to users on-demand. The cloud acts as a virtualized environment that can scale and allocate these resources dynamically based on user needs. In the present context, the edge cloud platform and the cloud platform may be implemented on separate servers (i.e., an edge server and a platform server) that are in remote communication with each other.
[0073] The processors of the servers or other computing devices discussed herein may be configured to perform, automatically, any and all functions, methods, and operations disclosed herein. The processors can, for example, perform operations embodying the function, method, or operation by, for example, executing code (e g., interpreting scripts) stored on memory (including any non-transitory computer readable medium). Accordingly, when a function is discussed of a functional unit is discussed it should be understood that one or more processors may be configured to provide such a function or functional unit by executing code stored in memory.
[0074] Returning to the example embodiment of FIG. 5, the example system is composed of two main parts. First, there are different site hosts, including multiple cell sites 440, one edge computing host 430, and one cloud platform 431 with the radio unit (RU), distributed unit (DU), and central unit (CU) components deployed for the different vBSs 431. Each radio unit (RU) is associated with one distributed unit (DU), and one distributed unit (DU) is associated with onecentral unit (CU). The different components are virtualized instances. Virtualization technologies enable to easily manage different instances automating their creation, life-cycle management and removal. Furthermore, they include mechanisms to allocate computing resources and migrate the components from one host to another. All the computing hosts have a metrics agent, which gathers different metrics from the components instances and the infrastructure. These metrics are saved into a metrics database 460. Second, embodiments of the present disclosure run in the non-real time (non-RT) ran intelligent controller (RIC) 470. The controller 470 can reach the edge cloud’s and the cloud platform's control plane.
[0075] For example, embodiments of the present disclosure can reach the following application processing interfaces (APIs):
[0076] An orchestration and management API: The orchestration and management API exists in the different hosts (e.g., edge cloud platform 430 and cloud platform 420), which allows for managing and orchestrating the deployed instances. The controller 470 can retrieve information from each container running into the platform. Also, the controller 470 can configure different computing resources, including the computing core set and the computing time for each container and the L3 cache memory.
[0077] An functional split API: The functional split API is an interface allowing the controller 470 to access the different distributed units (DUs) deployed and to configure the different distributed units (DUs) functional split with their corresponding central unit (CU).
[0078] The controller 470, according to an embodiment of the present disclosure, runs in discrete time intervals, which can be regarded as decision intervals. These decision intervals are from hundreds of milliseconds to seconds because embodiments of the present disclosure run in the non-RT RIC 470.
[0079] At the beginning of each decision interval, a context is input to the controller 470, which outputs the corresponding central unit (CU) placement 481, DU / CU functional split and computing configuration 482 to minimize the noisy neighbor effect at the edge computing platform. The controller 470 can fetch different metrics 480 from the metrics database 460, which stores all the measurement data collected 483 from the edge host computing platform's metrics agent(s), the cloud host cloud platform’s metrics agent(s), and / or the cloud cell site’s metrics agent(s).
[0080] The controller 470 learns the latent relationship between the context, actions, and computing consumption to optimize the system actions. The controller 470 learns how to minimize the noisy neighbor effect using a feedback signal, which captures the total computing consumption in the edge cloud 430 while satisfying the deployment requirements of the different instances.
[0081] In order to learn such latent relationships, and how to optimally configure the control units (CUs) and distributed units (DUs) to minimize the noisy neighbor effect, the controller 470 may implement a machine learning model. The model uses a set of metrics information (e g., collected from the distributed units (DUs)) to create a context vector for each virtual base station vBS 431. The context vector can be embedded into a fixed dimension space, and the controller 470 implements the model to learn the latent relationship between the context vectors and actions configurations (e g., actions for configuring the control units (CUs) and distributed units (DUs)). The model may include different artificial intelligence (Al) agents, which respectively map to different virtual base stations (vBSs) 431, and which learn how to collaborate in selecting the optimal configuration that minimizes the noisy neighbor effect. The behavior of the Al agents and / or the configuration of the model may be updated based on a feedback signal capturing use of computing resources for the present deployment of the vRAN.
[0082] In the following, the context, actions, and feedback signal are further described, as well as the controller's architecture.
[0083] The context input vector, according to an embodiment of the present disclosure, is based on the metrics collected for each distributed unit (DU) deployed into the system. These measurements are done in large time scales from seconds to minutes. For each distributed unit (DU), at least the following information is collected for each of its users:
[0084] Signal-to-noise ratio (SNR): The current SNR that they are experiencing in uplink and downlink. In the uplink, the distributed units (DUs) measure the SNR, whereas in the downlink this value can be inferred from the Channel Quality Indicator (CQI). The SNR may be used as part of the feedback signal and / or may be utilized as one of the metrics.
[0085] Traffic demand: Measurements are taken for the uplink and downlink traffic demands. This information can be retrieved from the distributed unit (DU) that is processing a user. The traffic demand may be used as part of the feedback signal and / or may be utilized as one of the metrics.
[0086] Aggregated traffic demand: The aggregated traffic demand the distributed unit (DU) is serving in both uplink and downlink. In a preferred embodiment, this metric is only retrieved when it is not possible to get per user (e.g., per mobile device) metrics for each distributed unit (DU). The aggregated traffic demand may be used as part of the feedback signal and / or utilized as one of the metrics.
[0087] Number of active users: The current number of active users the distributed unit (DU) is serving. In a preferred embodiment, this metric is only retrieved when it is not possible to get per user metrics for each distributed unit (DU).
[0088] The feedback signal may further include metrics on utilization of computing resources used by the distributed unit (DU) and / or the central unit (CU), utilization of computing resources provided by the deployment platform (e g., the edge server), and availability of additional computer resources.
[0089] The metrics agent stores all the previous information in the metrics database 460, which the controller 470 can access. Before inputting the data into its model, the controller 470 fetches the previous data and processes it to create the context input vector, which will serve as the input to the model.
[0090] Depending on which type of data that is available, the context input vector providing contextual information can be computed in the following way:
[0091] In a scenario where per-user metrics have been collected, the controller 470 may use each user's history samples to predict the probability distribution of the SNR and traffic demand for future decision intervals. For example, the controller 470 may use Kernel density estimation to estimate the probability density function from the samples. Furthermore, the different moments of the predicted probability distributions are computed, namely the mean, variance, skewness and / or kurtosis. A vector is constructed for each user, and the values of each vector are the moments for the previously-computed probability distributions. As there can be a different number of users in the cell, the different user vectors are embedded in a fixed dimension space. This can be done using a Relation Network, which takes as input all of the possible pairs of combinations of the vectors of the users, embeds each pair into a fixed dimension and aggregates all the embedded vectors. This vector can be used as the input context.
[0092] In a scenario where per-user metrics are not available, the controller 470 may use the aggregated metrics of the distributed unit (DU) to predict the probability distribution of the traffic demand and users for future decision intervals (e.g., using Kernel density estimation). As before, the different moments of the predicted probability distributions are computed, namely the mean, variance, skewness and kurtosis. Then, a vector is constructed whose values are the moments for the previously-computed probability distributions. This vector can be used as the input context.
[0093] The controller 470 can take different actions in the system depending on the deployment scenario and capabilities of the system.
[0094] In an embodiment, the actions that the controller 470 can take for each distributed unit (DU) and central unit (CU) pair include: (i) the placement of the central units (CUs); and (ii) computing configuration for the distributed units (DUs) deployed into the edge cloud host. If the central unit (CU) is placed co-located along with the distributed unit (DU) in the edge cloud 430, the computing configuration will be applied to the distributed unit (DU) and central unit(CU) as if they were one container. On the other hand, if the central unit (CU) is placed in the cloud platform 420, then the computing configuration of the distributed unit (DU) in the edge cloud 430 will only apply to itself
[0095] In an embodiment, the functional split configured between the distributed unit (DU) and the central unit (CU) can be additionally. The configuration options depend on the proprietary implementation being used. Configuring the functional split is generally useful in the scenario where both containers (e.g., distributed unit (DU) and central unit (CU)) are deployed separately. Changing the functional split configuration changes the workload of the distributed unit (DU) and the central unit (CU) because the computing demands are different since they are processing a different set of functions. Each functional split configuration sets the functions that the distributed unit (DU) and central unit (CU) will execute, thereby changing its workload. For example, if the split is changed so that the MAC and RLC layers are processed co-located with the central unit (CU), then the central unit (CU) has a different workload that changes its computational demand. This makes the processing much more demanding in the central unit (CU), while it releases some computing pressure away from the distributed unit (DU), which will have a less demanding workload. Thus, it is advantageous to configure the functional split along with the placement and the computing configuration when placing the central unit (CU) and provisioning the computing resources for both the distributed unit (DU) and the central unit (CU) according to the computing demand of their workloads. Thus, a joint configuration leads to less overprovisioning and reduced computing consumption.
[0096] The controller 470 uses a global feedback signal to optimize the edge cloud computing consumption. The less computing consumption, the less relevant the effect of the noisy neighbor problem. Whenever deciding on the distributed unit (DU) and central unit (CU) configuration, the aggregated computing consumption may be measured in the edge-computing host 430, and those measurements are used as the reward to the controller 470 for training. The Al agent is configured to minimize the aggregated computing consumption in the edge computing host 430 as a goal. Furthermore, the feedback signal takes into account whether the vBSs 431 are correctly functioning, checking that they are serving all users correctly and the cost of deploying central unit (CU) instances in the cloud platform 420. In this way, it is taken into account the deployment constraints of the different vBSs 431 and the cost of deploying central unit (CU) instances in the cloud platform 420 so that the minimization of the noisy neighbor problem is tuned with those constraints.
[0097] FIG. 5 shows the architecture of a controller 500, according to an embodiment of the present disclosure, using its different functional blocks. The controller 500 is composed of different agents 501 (e.g., Al agents), which map to a different vBS (e.g., vBSi, vBS2, . .. VBSN.)deployed into the system. The different agents 501 and a common system model 502 are trained using the feedback signal 503. This common model 502 is used for every agent 501 so that they learn how to collaborate to minimize the noisy neighbor effect. As different vBSs can have a different number of users, a context-embedding block 504 (e.g. using a Relation Network) is used at the input of each agent 501 so that variable context information 505 can be embedded for each vBS into a fixed-length vector.
[0098] The common model may be deployed using a multi-agent actor critic architecture. The agents that map to different DU-CU pairs can take as input the context of each DU-CU pair explained above and output the actions. The contexts and actions are input to the common model using the attention head to embed the information, and the common model outputs an estimation of the feedback signal. After this, the common model may be trained using the actual feedback signal, and the agents may also be trained using the actual feedback signal. In some implementations, an offline training phase may be used prior to online operation.
[0099] Each agent 501 will have different actions sets depending on whether they support changing the functional split between the distributed unit (DU) and the central unit (CU). However, the action set on the computing actions can be the same for every agent 501. The initial configuration of the actions sets may depend on the capabilities of the platform and the distributed units (DUs) and central units (CUs). Once all the agents 501 compute the different actions 506 for their respective vBS, the common model 502 trained with the feedback signal 503 is used by the different agents 501 to learn to collaborate to minimize the noisy neighbor problem.
[0100] An attention head 507 may be used to embed the different actions and contexts of the different agents 501 so that they learn how the other agents’ actions affects their actions. This improves convergence and teaches them how to collaborate. The attention head can be implemented using a neural network. The input to the attention head are the set of context and actions that the agents have chosen, and the output is fed directly to the common model.
[0101] The controller 500 will iteratively determine actions to take for each vBS for each decision interval, update each vBS based on the determined actions, and then update its model based on the measured results of the implemented action via the feedback signal 503.
[0102] While subject matter of the present disclosure has been illustrated and described in detail in the drawings and foregoing description, such illustration and description are to be considered illustrative or exemplary and not restrictive. Any statement made herein characterizing the invention is also to be considered illustrative or exemplary and not restrictive as the invention is defined by the claims. It will be understood that changes and modificationsmay be made, by those of ordinary skill in the art, within the scope of the following claims, which may include any combination of features from different embodiments described above.
[0103] The terms used in the claims should be construed to have the broadest reasonable interpretation consistent with the foregoing description. For example, the use of the article “a” or “the” in introducing an element should not be interpreted as being exclusive of a plurality of elements. Likewise, the recitation of “or” should be interpreted as being inclusive, such that the recitation of “A or B” is not exclusive of “A and B,” unless it is clear from the context or the foregoing description that only one of A and B is intended. Further, the recitation of “at least one of A, B and C” should be interpreted as one or more of a group of elements consisting of A, B and C, and should not be interpreted as requiring at least one of each of the listed elements A, B and C, regardless of whether A, B and C are related as categories or otherwise. Moreover, the recitation of “A, B and / or C” or “at least one of A, B or C” should be interpreted as including any singular entity from the listed elements, e g., A, any subset from the listed elements, e.g., A and B, or the entire list of elements A, B and C.
Claims
CLAIMSWhat is claimed is:
1. A computer-implemented method for configuring virtual base stations in a network comprising an edge server and a platform server, the edge server and the platform server hosting the virtual base stations, each respective virtual base station of the virtual base stations comprising a respective central unit and a respective distributed unit, the method comprising: for each of the virtual base stations, jointly determining a respective distributed unit computing configuration and a respective central unit placement configuration to minimize a noisy neighbor problem based on a set of metrics collected from the virtual base stations.
2. The method of claim 1, wherein for each of the virtual base stations, the joint optimization further j ointly optimizes a respective functional split configuration between the respective central unit and the respective distributed unit together with the distributed unit computing configuration and the central unit placement.
3. The method of any one of claims 1-2, the method further comprising, for each of the virtual base stations, commanding the respective virtual base station to update based on the respective distributed unit computing configuration, and / or the respective central unit placement configuration, and / or the respective functional split configuration.
4. The method of any one of claims 1-3, the method further comprising: generating a plurality of context vectors by generating, for each of the virtual base stations, a respective context vector based on metrics that are associated with the respective virtual base station and which are from the set of metrics collected; and embedding each of the context vectors into a fixed dimension space, wherein the joint optimization comprises using a machine learning model to learn latent relationships between the context vectors and actions configurations, and wherein, for each of the base stations, the respective actions configurations comprise actions for at least one of: placing the respective central unit at the edge server or the platform server, configuring computing resources for the respective distributed unit deployed at the edge server, or splitting functionalities between the respective distributed unit and the respective central unit.
5. The method of any one of claims 1-4, the method further comprising: using a machine learning model, learning a latent relationship between the context vectors and configurations of each of the virtual base stations currently deployed based on a feedback signal,wherein the feedback signal comprises information on a computing consumption used at the edge server and / or information on function of the virtual base stations.
6. The method of any one of claims 1-5, the method further comprising: for each of the virtual base stations, retrieving current metrics associated with the respective distribution unit and saving the current metrics as part of the set of metrics, wherein, for each of the virtual base stations, the current metrics comprise at least one of a signal-to-noise ratio, traffic demand, or number of active users, the metrics being associated with each user served by the respective distribution unit and / or being aggregated metrics for all of the users served by the respective distribution unit.
7. The method of any one of claims 1-6, wherein using a machine learning model comprises: executing a plurality of artificial intelligence (Al) agents to determine the respective actions configurations, each of the Al agents being associated with one of the virtual base stations.
8. The method of any one of claims 1-7, wherein the Al agents take as their inputs the context vectors.
9. The method of any one of claims 1-8, wherein using a machine learning model comprises: using a common model that is configured to learn how the Al agents collaborate to select the actions configurations to minimize the noisy neighbor problem, and updating the Al agents based on the common model.
10. The method of any one of claims 1-9, wherein the common model uses the feedback signal to learn the latent relationship between the context vectors and configurations of each of the virtual base stations and to provide rewards to the Al agents based on the feedback signal.
11. The method of any one of claims 1-10, wherein using a machine learning model comprises: using an attention head to embed different actions and contexts of the different Al agents, the agents being configured to learn how the other Al agents’ actions affects their actions based on the embeddings from the attention head.
12. The method of any one of claims 1-11, wherein the respective functional split configuration between the respective central unit and the respective distributed unit specifies which of the respective central unit or the respective distributed unit executes one or more of the following functions: medium access control (MAC) layer functions or radio link control (RLC) layer functions.
13. A computer system comprising one or more hardware processors which, alone or in combination, are configured to provide for execution of a method for configuring virtual base stations in a network comprising an edge server and a platform server, the edge server and the platform server hosting the virtual base stations, each respective virtual base station of the virtual base stations comprising a respective central unit and a respective distributed unit, the method comprising: for each of the virtual base stations, jointly determining a respective distributed unit computing configuration and a respective central unit placement configuration to minimize a noisy neighbor problem based on a set of metrics collected from the virtual base stations.
14. The computer system of claim 13, wherein the one or more hardware processors are configured to instantiate a machine learning model comprising: a plurality of artificial intelligence (Al) agents, each respective Al agent of the Al agents being configured to output actions for configuring one or more of: placement of the respective central unit at the edge server or the platform server, configuring computing resources for the respective distributed unit deployed at the edge server, or splitting functionalities between the respective distributed unit and the respective central unit; a common model that is configured to learn how the Al agents collaborate to select the actions to minimize the noisy neighbor problem, and to update the Al agents based on the common model; and an attention head that is configured to embed different actions and contexts of the different Al agents, wherein the Al agents are further configured to learn how the other Al agents’ actions affects their actions based on the embeddings received from the attention head, wherein the Al agents are each further configured to receive a respective context embedding as a respective input, the context embedding being generated by embedding a context vector into a fixed dimension space, the respective context vector being based on metrics that are associated with the respective virtual base station and which are from the set of metrics collected; and wherein each of the Al agents is configured to learn latent relationships between the respective context vector and the actions, and wherein for each of the virtual base stations, the actions output by the respective Al agent provides the jointly determined respective distributed unit computing configuration and respective central unit placement configuration that minimizes the noisy neighbor problem.
15. A tangible, non-transitory computer-readable medium having instructions thereon which, upon being executed by one or more hardware processors, alone or in combination, provide for execution of a method for configuring virtual base stations in a network comprising an edge server and a platform server, the edge server and the platform server hosting the virtual base stations, each respective virtual base station of the virtual base stations comprising a respective central unit and a respective distributed unit, the method comprising: for each of the virtual base stations, jointly determining a respective distributed unit computing configuration and a respective central unit placement configuration to minimize a noisy neighbor problem based on a set of metrics collected from the virtual base stations.
Citation Information
Patent Citations
Real-time radio access network analytics
WO2022245478A1
Ai-driven intelligent radio controller
WO2023030694A1
