WO2025250146A1

WO2025250146A1 - Artificial intelligence (ai) driven power savings in open radio access network (oran) heterogeneous network (hetnet) environment

Info

Publication number: WO2025250146A1
Application number: PCT/US2024/032146
Authority: WO
Inventors: Krishnan Venkataraghavan
Original assignee: Rakuten Mobile Inc; Rakuten Symphony Usa LLC
Current assignee: Rakuten Mobile Inc; Rakuten Symphony Usa LLC
Priority date: 2024-05-31
Filing date: 2024-05-31
Publication date: 2025-12-04
Anticipated expiration: 2026-11-30

Abstract

Artificial Intelligence (Al)-based power savings is provided in an Open Radio Access Network (ORAN) network. Metrics are obtained from servers in a network. N iterations are performed using the metrics to train a model of an artificial intelligence engine until an output of the model converges, First recommended processor core frequencies are generated using the trained model. The trained model and the first recommended processor core frequencies are provided to the servers. Model input and the first recommended processor core frequencies are received to create a reinforced model. Pre-implementation KPIs and Post-implementation KPIs are compared to identify a delta. Second recommended processor core frequencies are generated based on the delta, by performing M iterations until an output of the reinforced model converges to provide further optimized power savings and performance. The second recommended processor core frequencies are provided to the servers to provide the further optimized power savings and performance.

Description

ARTIFICIAL INTELLIGENCE (Al) DRIVEN POWER SAVINGS IN OPEN RADIO ACCESS NETWORK (ORAN) HETEROGENEOUS NETWORK (HETNET) ENVIRONMENT

FIELD

[0001] The present disclosure relates to Artificial Intelligence (Al) driven power savings in an Open Radio Access Network (ORAN) Heterogeneous Network (HETNET) environment.

BACKGROUND

[0002] The information disclosed in this background section is only for enhancement of understanding of the general background of the disclosure and should not be taken as an acknowledgement or any form of suggestion that this information forms the prior art already known to a person skilled in the art.

[0003] In a traditional Radio Access Network (RAN) system, the radio, hardware and software are proprietary. This means that nearly all of the equipment comes from one supplier and that operators are unable to, for example, deploy a network using radios from one vendor with hardware and software from another vendor. Such traditional RAN systems are implemented using purpose built hardware pre-optimized for RAN workloads.

[0004] Open RAN (ORAN) is a set of industry-wide standards for building mobile networks. ORAN is deployed on Commercially Off The Shelf (COTs) hardware. COTS hardware includes servers that include a CPU, memory, and storage components. COTS hardware is thus general purpose hardware that is not optimized for actual run workloads or traffic patterns. For example, ORAN has certain Key Performance Indicators (KPIs) that are to be met, such as a specific packet loss, throughput, and the like. Meeting the KPI goals ensures that an expected user experience is achieved. However, ORAN implemented with COTS are not initially configured for specific packet loss, latency, and the like. Thus, ORAN implemented with COTS is to be trained and optimized to the actual traffic.

[0005] ORAN architectures allow for disaggregation of hardware and software components of the RAN network, and the use of open interfaces and protocols to enable interoperability and integration of network elements from different vendors. Thus, communication service providers are able to use one supplier’s radios with another supplier’s RAN applications.

[0006] There are three primary elements in the RAN. A Radio Unit (RU) is where the radio frequency signals are transmitted, received, amplified and digitized. The RU is located near, or integrated into, the antenna. A Distributed Unit (DU) is where the real-time, baseband processing functions reside. The DU can be deployed at the cell site or concentrated in aggregated locations. A Centralized Unit (CU) is where the less time-sensitive packet processing functions typically reside.

[0007] One goal of ORAN is to support energy-saving within the RAN domain. Understanding the impact on power usage is important for cost savings and sustainability. A major part of energy consumption in mobile networks stems from the RAN. Electricity costs constitute a large percentage of cell site operating expenditures and a majority of energy cost is attributable to the ORAN and power amplifiers within a site. For example, energy spent on the radio network accounts for about 73% of the total power consumption or operator energy usage. The core network accounts for about 13%, the data center accounts for about 9%, and operations account for about 5% of the energy consumption.

[0008] However, current power saving technologies focus on a more reactive approach to power saving and are not optimized for close interaction with the ORAN framework where there are multiple different vendor combination of CU/DU/RU as well as hardware platforms.

SUMMARY

[0009] In at least embodiment, a method includes obtaining metrics from servers in a network. N iterations are performed using the metrics to train a mode of an artificial intelligence engine until an output of the model converges. First recommended processor core frequencies for the core processors of the servers are generated using the trained model to provide optimal power savings and performance for the core processors of the servers. The trained model and the first recommended processor core frequencies for the core processors are provided to the servers.

[0010] In at least one embodiment, a system is configured to obtain metrics from servers in a network. N iterations are performed using the metrics to train a mode of an artificial intelligence engine until an output of the model converges. First recommended processor core frequencies for the core processors of the servers are generated using the trained model to provide optimal power savings and performance for the core processors of the servers. The trained model and the first recommended processor core frequencies for the core processors are provided to the servers.

[0011] In at least one embodiment, a non-transitory computer-readable media having computer-readable instructions stored thereon, which when executed operations are performed for obtaining metrics from servers in a network. N iterations are performed using the metrics to train a model of an artificial intelligence engine until an output of the model converges. First recommended processor core frequencies for the core processors of the servers are generated using the trained model to provide optimal power savings and performance for the core processors of the servers. The trained model and the first recommended processor core frequencies for the core processors are provided to the servers.

BRIEF DESCRIPTION OF THE DRAWINGS

[0012] Features, aspects, and advantages of embodiments of the disclosure will be described below with reference to the accompanying drawings, in which like reference numerals denote like elements, and wherein:

[0013] Fig. 1 illustrates a mobile network according to at least one embodiment.

[0014] Fig. 2 is a block diagram of an Open Radio Access Network (O-RAN) according to at least one embodiment.

[0015] Figs. 3a-d show the breakdown of power consumption of a network, base station, and radio unit.

[0016] Fig. 4 illustrates a state diagram for the Power Performance States (P states) and Processor Idle Sleep States (C states) of CPUs according to at least one embodiment.

[0017] Fig. 5 illustrates the End to End Flow for performing Artificial Intelligence-Based Power Savings according to at least one embodiment.

[0018] Fig. 6 illustrates Stage 1 for training the Al engine according to at least one embodiment. [0019] Fig. 7 illustrates Stage 2 for Al-Based Power Saving according to at least one embodiment.

[0020] Fig. 8 illustrates overall Al-based power saving method according to at least one embodiment.

[0021] Fig. 9 illustrates a time chart of clock frequencies verses days according to at least one embodiment.

[0022] Fig. 10 shows the clustering for different clusters according to at least one embodiment. [0023] Fig. 11 is a flowchart of a method for Artificial Intelligence (Al) driven power savings in Open Radio Access Network (ORAN) Heterogeneous Network (HETNET) environment according to at least one embodiment.

[0024] Fig. 12 illustrates an embodiment of a device according to at least one embodiment. DETAILED DESCRIPTION

[0025] The following detailed description of example embodiments refers to the accompanying drawings. The present disclosure provides illustrations and descriptions, but is not intended to be exhaustive or to limit the implementations to the precise form disclosed. Modifications and variations are possible in light of the present disclosure or may be acquired from practice of the implementations. Further, one or more features or components of one embodiment may be incorporated into or combined with another embodiment (or one or more features of another embodiment). Additionally, the flowchart and description of operations provided below relate to at least one of the embodiments in the present disclosure. It should be noted that it is possible to make other embodiments that do not exactly match the flowchart and its description. It is understood that in other embodiments one or more operations may be omitted, one or more operations may be added, one or more operations may be performed simultaneously (at least in part).

[0026] It will be apparent that systems and/or methods, described herein, may be implemented in different forms of hardware, software, or a combination of hardware and software. The actual specialized control hardware or software code used to implement these systems and/or methods should not limit their implementations. Thus, the operation and behavior of the systems and/or methods are described herein without reference to specific software code. It is understood that software and hardware may be designed to implement the systems and/or methods based on the description herein.

[0027] Even though particular combinations of features are recited in the claims and/or disclosed in the specification, the particular combinations are not intended to limit the disclosure of implementations. In fact, many of these features may be combined in ways not specifically recited in the claims and/or disclosed in the specification. Even if a dependent claim directly depends on only one claim, the present disclosure may indicate that the dependent claim is dependent on other claims in the claim set.

[0028] No element, act, or instruction used herein should be construed as critical or essential unless explicitly described as such. Also, as used herein, the articles “a” and “an” (in other words, nouns not mentioned in the plural) are intended to include one or more items, and may be used interchangeably with “one or more.” Also, as used herein, the terms “has,” “have,” “having,” “include,” “including,” or the like are intended to be open-ended terms. Further, the phrase “based on” is intended to mean “based, at least in part, on” unless explicitly stated otherwise. Furthermore, expressions such as “at least one of [A] and [B],” “[A] and/or [B],” or “at least one of [A] or [B]” are to be understood as including only A, only B, or both A and B.

[0029] Further, spatially relative terms, such as “beneath,” “below,” “lower,” “above,” “upper” and the like, are used herein for ease of description to describe one element or feature’s relationship to another element(s) or feature(s) as illustrated in the figures. The spatially relative terms are intended to encompass different orientations of the device in use or operation in addition to the orientation depicted in the figures. The apparatus is otherwise oriented (rotated 90 degrees or at other orientations) and the spatially relative descriptors used herein likewise are interpreted accordingly.

[0030] The foregoing disclosure provides illustration and description but is not intended to be exhaustive or to limit the implementations to the precise form disclosed. Modifications and variations are possible in light of the above disclosure or may be acquired from practice of the implementations.

[0031] In at least one embodiment, a method includes obtaining metrics from servers in a network. N iterations are performed using the metrics to train a model of an artificial intelligence engine until an output of the model converges. First recommended processor core frequencies are generated for the core processors of the servers using the trained model to provide optimal power savings and performance for the core processors of the servers. The trained model and the first recommended processor core frequencies for the core processors are provided to the servers. The method further includes receiving model input and the first recommended processor core frequencies to create a reinforced model. Key Performance Indicators (KPIs) are received from pre-implementation of recommended processor core frequencies and KPIs based on implementation of the first recommended processor core frequencies are also received. The KPIs from the pre-implementation of the recommended processor core frequencies and the KPIs based on implementation of the first recommended processor core frequencies are compared to identify a delta. Based on the delta, performing M iterations until an output of the reinforced model converges to generate second recommended processor core frequencies to provide further optimized power savings and performance. The second recommended processor core frequencies are provided to the servers to provide the further optimized power savings and performance. [0032] Embodiments described herein provide method that provides one or more advantages. For example, Artificial Intelligence (Al)-based power savings is implemented to predict the optimal P/C states in which the RAN applications are able to operate to maintain a balance between performance and a power saving state. Based on the prediction of the optimal P and C states, savings in Capital Expenditures (CAPEX) and Operating Expenditures (OPEX) are provided for the operator.

[0033] Fig. 1 illustrates a mobile network 100 according to at least one embodiment.

[0034] In Fig. 1 , UE 1 (User Equipment 1 ) 110 and UE 2 1 12 access Mobile Network 100 via a Radio Access Network 120.

[0035] Radio Access Network 120 includes Radio Towers 121, 123, 125, and 127. Radio Towers 121, 123, 125, 127 are associated with RU (Radio Unit) 1 122, RU 2 124, RU 3 126, and RU 4 128, respectively.

[0036] RU 1 122, RU 2 124, RU 3 126, RU 4 128 handle the Digital Front End (DFE) and the parts of the PHY layer, as well as the digital beamforming functionality. RU 1 122 and RU 2 124 are associated with Distributed Unit (DU) 1 130, and RU 3 126 and RU 4 128 are associated with DU 2 132. DU 1 130 and DU 2 132 are responsible for real time Layer 1 and Layer 2 scheduling functions. For example, in 5G, Layer- 1 is the Physical Layer, Layer-2 includes the Media Access Control (MAC) , Radio link control (RLC), and Packet Data Convergence Protocol (PDCP) layers, and Layer-3 (Network Layer) is the Radio Resource Control (RRC) layer. Layer 2 is the data link or protocol layer that defines how data packets are encoded and decoded, and how data is to be transferred between adjacent network nodes. Layer 3 is the network routing layer and defines how data is moves across the physical network. [0037] DU 1 130 is coupled to the RU 1 122 and RU 2 124, and DU 2 132 is coupled to RU 3 126 and RU 4 128 . DU 1 130 and DU 2 132 run the RLC, MAC, and parts of the PHY layer. DU 1 130 and DU 2 132 include a subset of the eNB/gNB functions, depending on the functional split option, and operation of DU 1 130 and DU 2 132 are controlled by Centralized Unit (CU) 140. CU 140 is responsible for non-real time, higher L2 and L3. Server and relevant software for CU 140 is able to be hosted at a site or is able to be hosted in an edge cloud (datacenter or central office) depending on transport availability and the interface for the Fronthaul connections 150, 151, 153, 154. The server and relevant software of CU 140 is also able to be co-located at DU 1 130 or DU 2 132, or is able to be hosted in a regional cloud data center. [0038] CU 140 handles the RRC and PDCP layers. The gNB includes CU 140 and one or more DUs, e.g., DU 1 130, connected to CU 140 via Fs-C and Fs-U interfaces for a Control Plane (CP) 142 and User Plane (UP) 144, respectively. CU 140 with multiple DUs, e.g., DU 1 130, and DU 2 132, support multiple gNBs. The split architecture enables a 5G network to utilize different distribution of protocol stacks between CU 140, and DU 1 130 and DU 2 132, depending on network design and availability of the Midhaul 156. While two connections are shown between CU 140 and DU 1 130 and DU 2 132, CU 140 is able to implement additional connections to other DUs. CU 150, in 5G, is able to implement, for example, 256 endpoints or DUs. CU 140 supports the gNB functions such as transfer of user data, mobility control, RAN sharing (MORAN), positioning, session management, etc. However, one or more functions are able to be allocated to the DU. CU 140 controls the operation of DU 130 and DU 132 over the Midhaul interface 156.

[0039] Backhaul 158 connects the 4G/5G Core 160 to the CU 140. Core 160 may be, for example, up to 200 km away from the CU 140. Core 160 provides access to voice and data networks, such as Internet 170 and Public Switched Telephone Network (PSTN) 172.

[0040] RAN 120 is able to implement beamforming that allows for directional transmission or reception. 5G beamforming enables 5G connections to be more focused toward a receiving device. RAN 120 is also able to implement MIMO (Multiple Input Multiple Output), including mMIMO (massive MIMO), to provide an increases in throughput and signal-to-noise ratio (SNR). MIMO improves the radio link by using the multiple paths over which signals travel from the transmitter to the receiver. The multiple paths are de-correlated and this provides the opportunity to send multiple data streams over them.

[0041] According to at least one embodiment, a northbound platform for the network is provided, such as a Service Management and Orchestration (SMO)/NMS 180. SMO 180 oversees he orchestration aspects, and the management and automation of RAN elements. SMO 180 supports 01 , Al and 02 interfaces. Non-RT RIC (non-Real-Time RAN Intelligent Controller) 182 enables non-real-time control and optimization of RAN elements and resources, AI/ML workflow including model training and updates, and policy-based guidance of applications/features in Near-RT RIC 184. Near-RT RIC 184 enables near-real-time control and optimization of O-RAN elements and resources via fine-grained data collection and actions over the E2 interface. Near-RT RIC 184 includes interpretation and enforcement of policies from Non-RT RIC 182, and supports enrichment information to optimize control function. [0042] Near-RT RIC 184 obtains information associated with the beams that are passed to Non-RT RIC 182 and processed, for example, by an rApp at the Non-RT RIC 184, to generate an interference matrix. xApps are hosted on the Near-RT RIC 184 and are able to be used to optimize radio spectrum efficiency. rApps are specialized microservices operating on the Non- RT RIC 211. xApps and rApps provide control and management features and functionality. [0043] AI-Based Network Management is able to be provided at the 5G Edge via the rApps in the Non-RT RIC 182. Data is collected by a Node, such as an O-CU 140. Collected Data is processed. The ML Model at the Non-RT RIC 182 is Trained/Optimized using the processed data from the database. By implementing AI-Based Network Management at the 5G EDGE, performance is adjusted through continuous learning, and failures are handled by model monitoring.

[0044] While an O-RAN 120 is shown in Fig. 1 , embodiments described herein are applicable to O-RANs and Virtualized RANs (vRANs). O-RAN and vRAN disaggregate RAN hardware into three modules or functions, e.g., Radio Units (RUs) 122, 124, 126, 128, Distributed Units (DUs) 130, 132, and Centralized Units (CUs) 140. The software for these functions is decoupled from the purpose-built hardware and run on standardized, common off-the-shelf (COTS) hardware. O-RAN 120 further opens the software interfaces between radios and other network elements, whereas the interfaces between components in vRAN are still primarily based on closed or proprietary interfaces. A RAN Intelligent Controller (RIC) including Non- RT RIC 182 and RT RIC 184, is also able to be integrated with Multi-Access Edge Cloud (MEC) and vRAN. Herein, Radio Nodes refers to RUs 122, 124, 126, 128, Dus 130, 132, and CUs 140.

[0045] Fig. 2 is a block diagram of an Open Radio Access Network (O-RAN) 200 according to at least one embodiment.

[0046] In Fig. 2, Service Management and Orchestration (SMO) Framework 210 is an automation platform for Open RAN Radio Resources. SMO 210 oversees lifecycle management of network functions as well as O-Cloud. SMO 210 includes a Non-Real-Time (RT) Radio Access Network (RAN) Intelligent Controller (RIC) 211. SMO 210 also defines various SMO interfaces, such as the 01 216, 02 217, and Al 218 interfaces.

[0047] The Al interface 218 enables communication between the Non-RT RIC 211 and a Near-RT RIC 220 and supports policy management, data transfer, and machine learning management. The Al interface 218 is also used for policy guidance. SMO 210 provides fine- grained policy guidance such as getting User-Equipment to change frequency, and other data enrichments to RAN functions over the Al interface 218.

[0048] The 01 216 interface connects the SMO 210 to the RAN managed elements, which include the Near-RT RIC 220, O-RAN Centralized Unit (O-CU) 230, O-RAN Distributed Unit (0-DU) 240, and the Open Evolved NodeB (O-eNB) 260. The management and orchestration functions are received by the managed elements via the 01 interface 216. The SMO 210 in turn receives data from the managed elements via the 01 interface 216 for Al model training at the Non-RT RIC 21 1 . The 01 interface 216 is further used for managing the operation and maintenance (0AM) of multi-vendor Open RAN functions including fault, configuration, accounting, performance and security management, software management, and file management capabilities.

[0049] The 02 interface 217 is used to support cloud infrastructure management and deployment operations with O-Cloud 270 infrastructure that hosts the Open RAN functions in the network. The 02 interface 217 supports orchestration of O-Cloud infrastructure resource management (e.g., inventory, monitoring, provisioning, software management and lifecycle management) and deployment of the Open RAN network functions, providing logical services for managing the lifecycle of deployments that use cloud resources.

[0050] SMO 210 provides a common data collection platform for management of RAN data as well as mediation for the 01 216, 02 217, and Al 218 interfaces. Licensing, access control and AI/ML lifecycle management are supported by the SMO 210, together with legacy northbound interfaces. SMO 210 also supports existing Operational Support System (OSS) functions, such as service orchestration, inventory, topology and policy control.

[0051] SMO 210 also implements Federated Open Cloud Orchestration & Management (FOCOM) 214 and Network Function Orchestrator (NFO) 215. FOCOM 214 is responsible for managing the infrastructure (e.g., Clouds, Data centers, Clusters, Resources, etc.) on which the Network Slices, Services and Functions are deployed. The NFO 215 orchestrates the RAN network functions on top of them.

[0052] The Non-RT RIC 211 enables non-real-time (> 1 second) control of RAN elements and their resources through cloud-native microservice-based applications, which are referred to as rApps 212. An rApp 212 is able to implement an AI/ML Function 213. Non-RT RIC 211 communicates with applications called xApps 222 running on a Near-RT RIC 211 to provide policy-based guidance for edge control of RAN elements and their resources. The Non-RT RIC 211 provides non-real-time control and optimization of RAN elements and resources, AI/ML workflow, including model training of the AI/ML Function 213, updates, and policybased guidance of applications/features in Near-RT RIC 220.

[0053] Near-RT RIC 220 controls RAN infrastructure at the cloud edge. Near-RT RIC 220 controls RAN elements and their resources with optimization actions that typically take 10 milliseconds to one second to complete. The Near-RT RIC 220 receives policy guidance from the Non-RT RIC 211 and provides policy feedback to the Non-RT RIC 211 through the xApps 222.

[0054] The xApps 222 are used to enhance the RAN’s spectrum efficiency. The Near-RT RIC 220 manages a distributed collection of “southbound” RAN functions, and also provides “northbound” interfaces for operators: the 01 216 and Al 218 interfaces to the Non-RT RIC 211 for the management and optimization of the RAN. The Near-RT RIC 220 is thus able to self-optimize across different RAN types, like macros, Massive MIMO and small cells, maximizing network resource utilization for 5G network scaling.

[0055] Within the Near-RT RIC 220, the xApps 222 communicate via defined interface channels. An internal messaging infrastructure provides the framework to handle conflict mitigation, subscription management, app lifecycle management functions, and security. Data transfers are implemented via the E2 interface.

[0056] The 0-RAN is split into a Central Unit (CU) 230, a Distributed Unit (DU) 240, and a Radio Unit (RU) 250. The CU 230 is further split into two logical components, one for the Control Plane (CP) 232, and one for the User Plane (UP) 234. The logical split of the CU 230 into the CP 232 and UP 234 allows different functionalities to be deployed at different locations of the network, as well as on different hardware platforms. For example, CUs 230 and DUs 240 can be virtualized on servers at the edge, while the RUs 250 are able to be implemented on Field Programmable Gate Arrays (FPGAs) and Application- specific Integrated Circuits (ASICs) boards and deployed close to RF antennas.

[0057] The O-RAN Distributed Unit (0-DU) 240 is an edge server that includes baseband processing and radio frequency (RF) functions. The 0-DU 240 hosts radio link control (RLC), MAC, and a physical layer with network function virtualization or containers. O-DU 240 supports one or more cells, and the O-DUs are able to support one or more beams to provide the operating support for O-RU 250 by CUS (Control, User, and Synchronization) planes 252, and management (M) planes 254 through front-haul interfaces. [0058] The O-RU 250 processes radio frequencies received by the physical layer of the network. The processed radio frequencies are sent to the 0-DU 240 through FrontHaul (FH) interfaces 252, 254. The O-RU 250 hosts the lower PHY Layer Baseband Processing and RF Front End (RF FE), and is designed to support multiple 3GPP split options.

[0059] An Open-Evolved Node B (O-eNB) 260 provides the hardware aspect of the 0-RAN. The management and orchestration functions are received by the managed elements via the 01 interface 216. The SMO 210 in turn receives data from the managed elements via the 01 interface 216 for Al model training of AI/ML Functions 213 implemented by rApps 213 at Non-RT RIC 211. The O-eNB 260 communicates with the Near-RT RIC 220 via the E2 interface 224. E2 224 enables near-real-time loops through the streaming of telemetry from the RAN and the feedback with control from the Near-RT RIC 220. The E2 interface 224 connects the Near-RT RIC 220 with an E2 node, such as the O-CU-CP 232, O-CU-UP 234, the O-DU 240, and the O-eNB 260. An E2 node is connected to one Near-RT RIC 220, while Near-RT RIC 220 is able to be connected to multiple E2 nodes 224. The protocols over the E2 interface 224 are based on the control plane and supports services and functions of Near-RT RIC 220.

[0060] An Fl Interface 236 connects the O-CU-CP 232 and the O-CU-UP 234 to the 0-DU 240. Thus, the Fl interface 236 is broken into control and user plane subtypes and exchanges data about the frequency resource sharing and other network statuses. One O-CU 230 can communicate with multiple O-DUs 240 via Fl interfaces 236.

[0061] An El 238 interface connects the O-CU-CP 232 and the O-CU-UP 234. The El Interface 238 is used to transfer configuration data and capacity information between the O- CU-CP 232 and the O-CU-UP 234. The configuration data ensures the O-CU-CP 232 and the O-CU-UP 234 are able to interoperate. The capacity information is sent from the O-CU-UP 234 to the O-CU-CP 232 and includes the status of the O-CU-UP 234.

[0062] The O-DU 240 communicates with the O-RU 250 via an Open Fronthaul (FH) Control, User, and Synchronization (CUS) Plane Interface 252 and an M-Plane (Management Plane) Interface 254. As part of the CUS Plane Interface 252, the C-Plane (control plane) is a frame format that carries data in real-time control messages between the O-DU 240 and O-RU 250 for use to control user data scheduling, beamforming weight selection, numerology selection, etc. Control messages are sent separately for downlink (DL)- related commands and uplink (UL)-related commands. [0063] The U-Plane carries the user data messages between the O-DU 240 and O-RU 250, such as the in-phase and quadrature-phase (IQ) sample sequence of the orthogonal frequency division multiplexing (OFDM) signal. The S-plane includes synchronization messages used for timing synchronization between O-DU 240 and O-RU 250. The Control and User Plane is also used to send information specifying beamforming weights from the O-DU 240 to O-RU 250. Other information includes time resource and frequency resource information.

[0064] The M-Plane 254 connects the O-RU 250 to the O-DU 240, and an optional M-Plane 256 connects the O-RU 250 to the SMO 210. The O-DU 240 uses the M-Plane 254 to manage the O-RU 250, while the SMO 210 is able to provide FC APS (Fault, Configuration, Accounting, Performance, Security) services to the O-RU 250. The M-plane 254 supports the management features including startup installation, software management, configuration management, performance management, fault management and file management.

[0065] The M-Plane 254 is used by the O-DU 240 to retrieve the capabilities of the O-RU 250 and to send relevant configuration related to the C-Plane and U-Plane (data plane) to the O-RU 250. Together the 01 216 and Open-Fronthaul M-plane 254 interfaces provide a FCAPS interface with configuration, reconfiguration, registration, security, performance, monitoring aspects exchange with individual nodes, such as O-CU-CP 232, O-CU-UP 234, O-DU 240, and O-RU 250, as well as Non-RT RIC 220. O-Cloud 270 connects to Infrastructure Management Framework 280 via 02 Interface 217.

[0066] The O-Cloud 270 provides physical or logical infrastructure resources and performs workload management for O-RAN network functions. The O-Cloud 270 includes resource discovery and administration, network function provisioning, network function Fault, Configuration, Accounting, Performance, and Security (FCAPS), and software life cycle management. The O-Cloud 270 provides Infrastructure Management Services (IMS) 272 that communicates with the SMO 210.

[0067] The IMS 272 is responsible for physical resource allocation based on the request from the SMO 210 and resource tracking and management. The IMS 272 builds physical and logical inventories and shares them with the SMO 210 through the O2-M interface 217. The SMO 210 receives the inventory information from the IMS 272, updates its inventory accordingly, and makes a request to allocate a resource based on the inventory updates. The IMS 272 also provisions infrastructure resources and flexibly matches the resource demands of the O-RAN network functions. [0068] Non-RT RIC 211 collects from the O-Cloud 270 Fault, Configuration, Accounting, Performance, Security (FCAPS) data over the 02 interfaces, and collects data from E2 node over the 01 interface. An Artificial Intelligence/Machine Learning (AI/ML) model 213 is trained and deployed to generate recommendations for power savings according to at least one embodiment. The O-Cloud 270 receives recommendations from SMO 210 for optimizing energy consumptions for various resources of the O-Cloud 270, and generates recommendations for energy saving. A Management Platform 280 is used to control the Artificial Intelligence (Al) Driven Power Savings according to at least one embodiment.

[0069] Figs. 3a-d show the breakdown of power consumption of a network, base station, and radio unit.

[0070] In Fig. 3 a, operator energy use breakdown 310 is shown. The Radio Access Network (RAN) 312 is shown to be responsible for about 73% of total energy consumption. RAN 312 includes several cell sites with site infrastructure equipment and base station equipment. The Core Network 314 is responsible for about 13% of energy consumption. The Core Network 314 includes the large-capacity core routers and high-speed fiber optic cables which connect consumers and businesses to Data Centers. The Data Center 316 is responsible for about 9% of energy consumption. The Data Center 316 includes servers, power supplies, cooling systems. Finally, other operations 318 are responsible for about 5% of energy consumption. [0071] Fig. 3b shows the power consumption breakdown of a Radio Unit (RU) 330. The Power Amplifier 332 is responsible for about 59% of energy consumption of a RU. The Power Amplifier 332 is used to amplify signals for transmission from the base station to mobile devices. The Analog Front End (AFE) and Digital Front End (DFE) 334 are responsible for about 35% of energy consumption of a RU. The AFE connects to the radiating panel and contains analog components like power amplifiers, filters, drivers and baluns and may contain switches and circulators. The AFE amplifies the Transmit (Tx) and Receive (Rx) signals to and from the antennas. The AFE provides dynamic range for the Rx and Tx paths, isolates the paths, and manages any noise introduced by the power amplifier stages. The DFE prepares and multiplexes the signals created by the baseband processing subsystem and sends them to the RF power amplifier for transmission. The Power Supply 336 is responsible for about 6% of energy consumption of a RU.

[0072] Fig. 3c shows the power consumption breakdown of a Base Station (BS) with air conditioning 350. The Remote Radio Head (RRH) 352 is responsible for about 40% of energy consumption of a BS with air conditioning. A RRH 352 is a remote radio transceiver that connects to radio base station unit via electrical or wireless interface. The RRH 352 is termed “Remote” as it is usually installed on a mast-top, or tower-top location that is physically some distance away from the base station hardware which is often mounted in an indoor rackmounted location The Baseband Processing 354 is responsible for about 6% of energy consumption of a BS with air conditioning. Baseband Processing 354 includes LI , L2, L3, and transport processing. The Main Control 356 is responsible for about 4% of energy consumption of a cooled BS. The Main Control 356 manages and controls the base station system, overseeing functions such as data processing, signaling, and resource management. The Cooling System 358 is responsible for about 40% of energy consumption of a cooled BS. The Power Supply 360 is responsible for about 7% of energy consumption of a cooled BS. Other Components 362 are responsible for about 3% of energy consumption of a cooled BS.

[0073] Fig. 3d shows the power consumption breakdown of a Base Station (BS) without air conditioning 370. The Remote Radio Head 372 is responsible for about 67% of energy consumption of a BS without air conditioning. The Baseband Processing 374 is responsible for about 10% of energy consumption of a BS without air conditioning. The Main Control 376 is responsible for about 7% of energy consumption of an uncooled BS. The Power Supply 378 is responsible for about 11% of energy consumption of an uncooled BS. Other Components 380 are responsible for about 5% of energy consumption of an uncooled BS.

[0074] Existing power savings solutions involve reactive power saving techniques that are based on traffic patterns. For example, Micro-Sleep Transmit (TX) is one type of power saving technique. Micro-Sleep TX involves switching from 32-transmit/receive (32TR) antennas to 4-transmit/receive antennas. These solutions work reactively and there is a need to further optimize this mechanism to consider the RAN workload being time sensitive and further the need to maintain Key Performance Indicators (KPIs) including, but not limited to, delay, jitter, latency, user experience, throughput, packet loss, and the like. Microsleep TX depends on the RU (Radio Unit ) transition times to transition between active and inactive states. These transition times are able to be as high as 500 milliseconds (ms). This results in packet loss/lower latency as well loss of network traffic. For example, RAN workloads are time sensitive and have a one millisecond upper limit. Thus, a transition that takes even 100 milliseconds results in packet loss.

[0075] Resource reconfiguration techniques for power saving depend on the traffic payloads. A decision is made whether to maintain 4 transmit 4 receive (4t4r) antennas towards the user or to transition to 2 transmit 4 receive (2t4r) towards the user. Resource reconfiguration techniques result in lower spectral efficiencies as well loss of transmit diversity combining gains in the Downlink (DL) resulting in poorer user perception and subsequent impairments in operator chum and/or user retention.

[0076] Adaptive Cell switch-off is based on historic traffic patterns that are extrapolated to predict cell state. However, this is more optimal for a binary decision of Off/On based on periods of traffic and periods of no traffic. The coverage pattern is being switched on and off. This results in longer transition times and is less optimal for bursty traffic patterns which are common in a RAN deployment.

[0077] These techniques are reactive techniques. Also, these techniques are not optimized to consider the RAN workloads because RAN workloads are time sensitive. Time sensitive scheduling on the RAN happens at an order of 0.5 milliseconds in 5G and one millisecond in Long-Term Evolution (LTE) . In response to the delay between the actual RAN traffic and the actual cell state being more than the one millisecond, further degradations occur and KPIs are not able to be maintained, such as delay, jitter, latency and overall user experience such throughput and packet loss. Hence there is a need to not only optimize the existing solution but further extend that to ORAN deployments which works on general purpose hardware rather than on purpose-built hardware .

[0078] Fig. 4 illustrates a state diagram 400 for the Power Performance States (P states) and Processor Idle Sleep States (C states) of Central Processing Units (CPUs), Graphics Processor Units (GPUs), and the like, according to at least one embodiment. Herein, processor is used generically to refer to any type of CPU, GPU, and the like.

[0079] Processor power management technologies are defined the Unified Extensible Firmware Interface (UEFI) forum in the Advanced Configuration and Power Interface (ACPI) specification. The ACPI specification is an interface specification that includes both software and hardware elements. The ACPI specification and are divided into two categories or states: Power Performance States (ACPI P states) and Processor Idle Sleep States (ACPI C states).

[0080] P-States provide a way to scale the frequency and voltage at which the processor core runs so as to reduce the power consumption of the processor. The number of available P-States can be different for each model of processor, even those from the same family.

[0081] C-States are states when the processor has reduced or turned off selected functions. Different processors support different numbers of C-states in which various parts of the processor are turned off. Generally, higher C-states turn off more parts of the processor, which significantly reduces power consumption. Processors are able to have deeper C-states that are not exposed to the operating system.

[0082] Processors have multiple functions, such as processor core pinning, processor clock frequency, and the like. Processor core pinning involves binding a process or a thread to a specific processor core, and preventing the process from running on any other core. In transitions between C state 1 to C state 2, certain functions of the CPU are turned off. In response to transitioning from P0 to Pl, for example, the frequency and voltage at which the processor is operating at are reduced.

[0083] In Fig. 4, a processor is able to be in an Idle State 410. The processor is able to transition to a Power Performance State. In Fig. 4, the processor is able to transition to a P0 state 420, wherein the processor operates at a frequency of 2.6 Gigahertz (GHz). The processor is also able to transition to a Pl state 430, wherein the processor operates at a frequency of 2.0 GHz, or to a P2 state 440, wherein the processor operates at a frequency of 1.6 GHz. As shown in Fig. 4, the processor is able to transition from a current state to any of the other states. Those skilled in the art understand that the frequencies for P0 420, Pl 430, and P2 440 are provided as examples and other frequencies are able to be implemented without departing from the embodiments described herein.

[0084] Fig. 5 illustrates the End to End Flow 500 for performing Artificial Intelligence-Based Power Savings according to at least one embodiment.

[0085] In Fig. 5, metrics are extracted from servers 510. The metrics are able to be extracted in real time. For example, a Telegraf agent is able to collect, process, aggregate, and write metrics, logs, and other data. IPMI (Intelligent Platform Management Interface) is a set of standardized specifications for hardware-based platform management systems that makes it possible to control and monitor servers centrally. Also, Foresight is able to be used to track, analyze and display key performance indicators (KPI), metrics and data points to monitor servers in the network.

[0086] Unsupervised Learning is performed 520 using the extracted metrics. From the extracted metrics, N iterations using the metrics to train a model of an artificial intelligence engine until an output of the model converges. The Al engine is able to learn, for example, workload signatures, network traffic patterns, time of day patters, or Operating System (OS) and Kernel statistics. Other parameters or information is also able to be learned.

[0087] The Al engine predicts processor loads and produces application specific predicted P and C states 530, such as the recommended core level processor frequencies. Application specific predicted data includes data on a particular run, a core or Infrastructure Management Services (IMS) application.

[0088] The Artificial Intelligence (Al) Driven Power Savings process according to at least one embodiment includes two stages. Stage 1 involves providing metrics to a Central Data Server for training the Al engine.

[0089] Fig. 6 illustrates Stage 1 for training the Al engine 600 according to at least one embodiment.

[0090] In Fig. 6, model input 610 is provided to the Central Data Server 620 for training an Al engine 622. Metrics are able to include CPU Frequency, Fan Speed, Overheating Metrics (alarms), CPU Pinning, and the like 630. The metrics also are able to include Traffic Patterns, Packet Loss, Throughput, UE Connection Loss metrics 640. Further metrics include Time Patterns 650 and data regarding the OS, Kernel, Application Help Charts, Virtual Infrastructure Management (VIM), Kubemetes Metrics, and the like 660. A person skilled in the art understands that the listing of data for the model is provided as an example and that other or different data is able to be used.

[0091] Processor cores in servers in the network are operated at a particular clock frequency, fan speed, voltage, and the like. The timestamped metrics for these patterns are provided to the Central Data Server. The Al 622 performs workload clustering 670 to generate a model. Servers are segmented through unsupervised learning based on their workload signatures, e.g., CPU/IPMFNetwork/Input/Output (IO). The Al 622 predicts processor core workloads based on the metrics. Processor cores are clustered into patterns based, for example, on a particular traffic pattern, throughput, User Equipment (UE) Connection Lost according to a particular time of a day. Models are built for each cluster and processor throttling is recommended on a core level. Based on the clustering, the Al predicts 622 the optimal P/C States 680 (e.g., core level processor frequencies) at which those clusters are able to operate to produce an optimal power savings and performance. The pre trained model and recommended processor core frequencies are provided by Central Data Server to the servers where the actual applications are running. [0092] Fig. 7 illustrates Stage 2 for AI-Based Power Saving 700 according to at least one embodiment.

[0093] In Fig. 7, Model Input from Stage 1 710 is used to create a reinforced model for Stage 2. The KPIs from the pre-implementation of Stage 1 recommended processor core frequencies and KPIs from post-implementation of the Stage 1 recommended processor core frequencies 720 are provided to the Stage 2 model 730. The KPIs from the pre-implementation of from Stage 1 recommended processor core frequencies are compared to the KPIs from postimplementation of the Stage 1 recommended processor core frequencies 720 to determine whether the KPIs are following the expectations after the Stage 1 recommended processor core frequencies have been implemented. P and C State recommendations from Stage 1 740 are provided to mitigate the delta between KPIs from the pre-implementation of Stage 1 recommended processor core frequencies and the KPIs from the pre-implementation of from Stage 1 recommended processor core frequencies. Thus, Inputs of Stage 1 are used for a reinforcement learning so that the optimized P and C States are fed to the servers so that the delta between KPIs from the pre-implementation of Stage 1 recommended processor core frequencies and the KPIs from the pre-implementation of from Stage 1 recommended processor core frequencies is mitigated. This ensures that the KPIs or user experience remains intact between the pre and the post implementations of recommended P and C states. In response, Stage 2 produces new P States 750.

[0094] Fig. 8 illustrates overall Al-based power saving method 800 according to at least one embodiment.

[0095] Fig. 8 shows the Network Orchestrator Touchpoints (Northbound/Southbound) 802 for implementing the Al-based power saving method 800 according to at least one embodiment. In Fig. 8, the ORAN 810 is shown. ORAN 810 is describe above with respect to Fig. 2 and is not further described here. Multiple cell sites in Network 820 are shown. Multiple cell sites in Network 820 and ORAN 810 are deployed using Commercially Off The Shelf (COTs) hardware. COTS hardware includes servers that include a processor, memory, and storage components. Raw data is collected either directly from the network devices by the Orchestrator 830 or through Environment Monitoring Systems (EMS) 832 via the Observability Framework 840 as part of ORAN 810. Operating System (OS) and Kernel statistics 834 are also collected. An Al engine is provided by Central Data Server 850. The Al engine provides Workload Clustering. Data, e.g., gathered set of metrics, is first used for Model Training and Inference at Stage 1 860. The Pre-Trained Model 870 is then provided to a Data Center 880 in Stage 2 890. As described above, inputs of Stage 1 860 are used for a reinforcement learning in Stage 2 890 so that the optimized P and C States are fed 882 from Data Center 880 to the servers in the Network 820 so that the delta between KPIs from the pre-implementation of Stage 1 recommended processor core frequencies 860 and the KPIs from the pre-implementation of from Stage 1 recommended processor core frequencies is mitigated in Stage 2 890. This ensures that the KPIs or user experience remains intact between the pre-implementation of recommended processor core frequencies of Stage 1 860 and the post-implementations of Stage 1 recommended processor core frequencies in response to generating in Stage 2 890 of recommended P and C states. In response, Stage 2 890 produces new P States that are provided to the servers in the Network 820 to provide optimized energy savings and performance.

[0096] Fig. 9 illustrates a time chart 900 of clock frequencies verses days according to at least one embodiment.

[0097] In Fig. 9, the Y axis 910 represents clock frequency and the X axis 920 is the time period, e.g., days. A person skilled in the art recognizes that the Y axis 910 is not meant to be limited to the illustrated clock frequencies, and X axis 920 is not meant to be limited to days. For example, other clock frequencies are able to be represented. Also the Y axis 910 is able to represent months, days/hours/seconds and the like as desired. At the bottom of Fig. 9, traffic patterns 930 are shown. Also shown is a prediction graph 940 that is generated based on the traffic patterns to optimize power saving in the network. As shown, the traffic patterns 930 show peaks and troughs. Other metrics 950, 960, for example, represent model input. For example, other metrics represented by other graphs, such as 950, 960, are able to include Processor Core Frequency, Fan Speed, Overheating Metrics (alarms), Processor Core Pinning, and the like, Packet Loss, Throughput, UE Connection Loss metrics and the like, Time Patterns and data regarding the OS, Kernel, Application Helm Charts, Virtual Infrastructure Management (VIM), Kubemetes Metrics, and the like.

[0098] The generation of the Prediction graph is based on a peak in the traffic patterns 970 corresponding to a peak in the predictions graph 972 . Similarly, a trough in the traffic pattern 980 should correspond to a trough in the predictions graph 982. As seen in Fig. 9, a peak in traffic pattern 970 is shown to correspond to a peak in the predictions graph 972, and a trough in the traffic pattern 980 is shown to correspond to a trough in the predictions graph 982. In Fig. 9, the other graphs 950, 960, for example, represent other metrics that are also able to be considered to generate prediction graph 940.

[0099] Fig. 10 shows the clustering for different clusters 1000 according to at least one embodiment.

[0100] In Fig. 10, cluster 6 1010 represents a cluster with the same workload signature 1012. The cluster for cluster 6 1010 will use the same model 1014, e.g., model 1. However, cluster 1 1020 is able to use a different model 1022, e.g., model 2. The P and C States for a processor core of cluster 6 1010 used by model 1 1014 are able to be different than the P and C States for a processor core of cluster 1 1020 used by model 2 1022. Fig. 10 also shows clusters 0, cluster 3, cluster 4, cluster 5, and cluster 7. However, more or less clusters are able to be considered. [0101] As described above, servers are segmented through unsupervised learning based on their workload signatures, e.g., CPU/IPMFNetwork/Input/Output (IO). Models are built for each cluster and processor throttling is recommended on a core level. Applications are able to be run on a particular core. Through processor core pinning of workloads, certain applications use certain cores of the processor. A processor includes multiple cores, e.g., 10 cores, 20 cores, depending on the model of the server.

[0102] A particular recommended processor core frequency cannot be applied to the processor itself. The recommended processor core frequency depends on which processor core is being targeted. Thus, in response to a processor having 20 cores, the model is able to provide 20 different frequency outputs.

[0103] Fig. 11 is a flowchart 1100 of a method for Artificial Intelligence (Al) driven power savings in an Open Radio Access Network (ORAN) Heterogeneous Network (HETNET) environment according to at least one embodiment.

[0104] In Fig. 11, the process begins SI 102 and metrics are obtained from servers in a network S 1110. Referring to Fig. 5, metrics are extracted from servers 510. The metrics are able to be extracted in real time. For example, a Telegraf agent is able to collect, process, aggregate, and write metrics, logs, and other data. IPMI (Intelligent Platform Management Interface) is a set of standardized specifications for hardware -based platform management systems that makes it possible to control and monitor servers centrally. Also, Foresight is able to be used to track, analyze and display key performance indicators (KPI), metrics and data points to monitor servers in the network. Referring to Fig. 6, model input 610 is provided to the Central Data Server 620 for training an Al engine 622. Referring to Fig. 8, raw data is collected either directly from the network devices by the Orchestrator 830 or through Environment Monitoring Systems (EMS) 832 via the Observability Framework 840 as part of ORAN 810. Operating System (OS) and Kernel statistics 834 are also collected.

[0105] A model of an artificial intelligence engine is trained by performing N iterations using the metrics SI 114. Referring to Fig. 5, Unsupervised Learning is performed 520 using the extracted metrics. From the extracted metrics, the Al engine performs N iterations using the metrics to train a model of an artificial intelligence engine using the metrics until an output of the model converges. The Al engine is able to learn, for example, workload signatures, network traffic patterns, time of day patters, or Operating System (OS) and Kernel statistics. Other parameters or information is also able to be learned. The Al engine predicts processor loads and produces application specific predicted P and C states 530, such as the recommended core level processor frequencies. Application specific predicted data includes data on a particular run, a core, or IMS application. Referring to Fig. 8, an Al engine is provided by Central Data Server 850. The Al engine provides Workload Clustering. The data is first used for Model Training and Inference at Stage 1 860.

[0106] First recommended processor core frequencies are generated for the processor cores of the servers using the trained model to provide the optimal power savings and performance for the processor cores of the servers SI 122. Referring to Fig. 6, based on the clustering, the Al predicts the optimal P/C State 680 (e.g., core level processor frequencies) at which those clusters are able to operate to produce an optimal power savings and performance.

[0107] The trained model and the first recommended processor core frequencies for the processors are provided to the servers SI 126. Referring to Fig. 6, the pre trained model and recommended processor core frequencies are shipped from the Central Data Server to the servers where the actual application is running. Referring to Fig. 8, the Pre-Trained Model 870 is then provided to a Data Center 880 in Stage 2 890.

[0108] Model input and the first recommended processor core frequencies are received for creating a reinforced model SI 130. Referring to Fig. 7, model Input from Stage 1 710 is used to create a reinforced model for Stage 2. Referring to Fig. 8, inputs of Stage 1 860 are used for a reinforcement learning in Stage 2 890 so that the optimized P and C States are fed 882 from Data Center 880 to the servers in the Network 820

[0109] KPIs from pre-implementation of recommended processor core frequencies and KPIs based on implementation of the first recommended processor core frequencies are received SI 134. Referring to Fig. 7, the KPIs from the pre-implementation of Stage 1 recommended processor core frequencies and KPIs 720 from post- implementation of the Stage 1 recommended processor core frequencies are provided to the Stage 2 model 730.

[0110] The KPIs from the pre-implementation of recommended processor core frequencies and KPIs from the post-implementation of the first recommended processor core frequencies are compared to identify a delta SI 138. Referring to Fig. 7, the KPIs from the pre-implementation of Stage 1 recommended processor core frequencies are compared to KPIs from postimplementation of the Stage 1 recommended processor core frequencies 720 to determine whether the KPIs are following the expectations after the Stage 1 Stage 1 recommended processor core frequencies have been implemented. Thus, Inputs of Stage 1 are used for a reinforcement learning so that the optimized P and C States fed to the servers so that the delta between KPIs from the pre-implementation of Stage 1 and the KPIs from the preimplementation of from Stage 1 recommended processor core frequencies is mitigated. Referring to Fig. 8, inputs of Stage 1 860 are used for a reinforcement learning in Stage 2 890 so that the optimized P and C States are fed 882 from Data Center 880 to the servers in the Network 820 so that the delta between KPIs from the pre-implementation of Stage 1 recommended processor core frequencies 860 and the KPIs from the pre-implementation of from Stage 1 recommended processor core frequencies is mitigated in Stage 2 890 is mitigated. This ensures that the KPIs or user experience remains intact between the pre-implementation of recommended processor core frequencies of Stage 1 860 and the post-implementations of Stage 1 recommended processor core frequencies in response to generating in in Stage 2 890 of recommended P and C states.

[0111] Based on the delta, M iterations are performed until an output of the reinforced model converges to generate second recommended processor core frequencies to provide further optimized power savings and performance SI 142. Referring to Fig. 7, inputs of Stage 1 are used for a reinforcement learning so that the optimized P and C States are fed to the servers so that the delta between KPIs from the pre-implementation of Stage 1 and recommended processor core frequencies and the KPIs from the post-implementation of from Stage 1 recommended processor core frequencies is mitigated. This ensures that the KPIs or user experience remains intact between the pre and the post implementations of recommended P and C states. In response, Stage 2 produces new P States 750.

[0112] The second recommended processor core frequencies are provided to the servers to provide the further optimized power savings and performance S 1146. Referring to Fig. 8, Stage 2 890 produces new P States that are provided to the servers in the Network 820 to provide optimized energy savings and performance.

[0113] The process then terminates SI 150.

[0114] According to at least one embodiment, the method includes obtaining metrics from servers in a network and performing N iterations using the metrics to train a model of an artificial intelligence engine until an output of the model converges. First recommended processor core frequencies are generated for the core processors of the servers using the trained model to provide an optimal power savings and performance for the core processors of the servers. The trained model and the first recommended processor core frequencies for the core processors of the servers are provided to the servers. The method further includes receiving model input and the first recommended processor core frequencies to create a reinforced model. Key Performance Indicators (KPIs) from pre-implementation of the first recommended processor core frequencies and KPIs based on implementation of the first recommended processor core frequencies are received. The KPIs from pre-implementation of the first recommended processor core frequencies and the KPIs based on implementation of the first recommended processor core frequencies are compared to identify a delta. Based on the delta, M iterations are performed until an output of the reinforced model converges to generate second recommended processor core frequencies to provide further optimized power savings and performance. The second recommended processor core frequencies are provided to the servers to provide the further optimized power savings and performance.

[0115] Fig. 12 illustrates an embodiment of a device 1200.

[0116] As shown in Fig. 12, the device 1200 includes processor 1210, a memory 1220, a storage component 1230, an input component 1240, an output component 1250, a communication interface 1260, and a bus 1270.

[0117] The processor 1210, as used herein, means any type of computational circuit that may comprise hardware elements and software elements. The processor 1210 is able to be embodied as a multi-core processor, a single core processor, or a combination of one or more multi-core processors and/or one or more single core processors, a distributed processing system, or the like. The processor 1210 may be a Central Processing Unit (CPU), a graphics processing unit (GPU), an accelerated processing unit (APU), an application-specific integrated circuit (ASIC), or another type of processing component.

[0118] Memory 1220 includes a non-transitory computer readable medium. Memory 1220 includes a random-access memory (RAM), a read only memory (ROM), and/or another type of dynamic or static storage device (e.g., a flash memory, a magnetic memory, and/or an optical memory) that stores information and/or instructions for use by processor 1210. The memory 1220 comprises machine-readable instructions which are executable by the processor 1210. These machine-readable instructions when executed by the processor 1210 cause the processor 1210 to perform one or more method steps of an embodiment as described above.

[0119] Storage component 1230 stores information and/or software related to the operation and use of the device 1200. For example, storage component 1230 may include a hard disk (e.g., a magnetic disk, an optical disk, a magneto-optic disk, and/or a solid-state disk), a compact disc (CD), a digital versatile disc (DVD), a floppy disk, a cartridge, a magnetic tape, and/or another type of non-transitory computer-readable medium, along with a corresponding drive.

[0120] Input component 1240 is configured to receive information, such as user input. For example, the input component 1240 may include, but not be limited to, a touch screen display, a keyboard, a keypad, a mouse, a button, a switch, and/or a microphone. Additionally, or alternatively, the input component 1240 may include a sensor for sensing information (e.g., a global positioning system (GPS), an accelerometer, a gyroscope, and/or an actuator).

[0121] Output component 1250 is configured to provide output information from the device 1200. For example, the output component 1250 may be, but not limited to, a display, a speaker, an instruction device to an external device, and/or one or more light-emitting diodes (LEDs).

[0122] Communication interface 1260 is an interface that provides a communication connection to other devices, such as external devices and internal devices. The connection by the communication interface 1260 can be a wired connection, a wireless connection, or a combination of wired and wireless connections, and can be a direct connection or an indirect connection via a communication network that exists between the device 1200 and other devices. In other words, the communication interface 1260 is not limited.

[0123] The bus 1270 acts as an interconnect between the processor 1210, the memory 1220, the storage component 1230, the input component 1240, the output component 1250, and the communication interface 1260 of the device 1200. The bus 1270 may include a wired interconnection or a wireless interconnection.

[0124] The number and arrangement of components shown in FIG. 12 are provided as an example. In practice, device 1200 may include additional components, fewer components, different components, or differently arranged components than those shown in FIG. 12. Additionally, or alternatively, a set of components (e.g., one or more components) of device 1200 may perform one or more functions described as being performed by another set of components of device 1200. Further, one or more method steps described in any of the embodiments may be performed utilizing a plurality of devices 1200 in communication with one another.

[0125] Embodiments described herein provide method that provides one or more advantages. For example, Artificial Intelligence (Al)-based power savings is implemented to predict the optimal P/C states in which the RAN applications are able to operate to maintain a balance between performance and a power saving state. Based on the prediction of the optimal P and C states, savings in Capital Expenditures (CAPEX) and Operating Expenditures (OPEX) are provided for the operator.

[0126] [1] An aspect of this description is directed to a method that includes obtaining metrics from servers in a network, performing N iterations using the metrics to train a model of an artificial intelligence engine until an output of the model converges, generating first recommended processor core frequencies for the core processors of the servers using the trained model to provide an optimal power savings and performance for the core processors of the servers, and providing the trained model and the first recommended processor core frequencies for the core processors to the servers.

[0127] [2] The method described in [1], wherein the generating the first recommended processor core frequencies are based on an analysis of the metrics to identify information including one or more of workload signatures, network traffic patterns, time of day patters, or Operating System (OS) and Kernel statistics.

[0128] [3] The method described in any of [1] to [2], further includes receiving model input and the first recommended processor core frequencies to create a reinforced model, receiving Key Performance Indicators (KPIs) from pre-implementation of the first recommended processor core frequencies and KPIs based on implementation of the first recommended processor core frequencies, comparing the KPIs from pre-implementation of the recommended processor core frequencies and KPIs based on implementation of the first recommended processor core frequencies to identify a delta, based on the delta, performing M iterations until an output of the reinforced model converges to generate second recommended processor core frequencies to provide further optimized power savings and performance, and providing the second recommended processor core frequencies to the servers to provide the further optimized power savings and performance.

[0129] [4] The method described in [3], wherein the generating the second recommended processor core frequencies includes segmenting one or more clusters of the servers using unsupervised learning and determining the second recommended processor core frequencies for the one or more clusters to produce the optimal power savings and performance for each of the one or more clusters, wherein the determining the second recommended processor core frequencies for the one or more clusters includes determining optimized P states and C state for core processors of the one or more clusters of the servers and choosing a combination of the P states and C states to produce the optimal power savings and performance for the core processors of each of the one or more clusters of the servers.

[0130] [5] The method described in any of [3] to [4], wherein the generating the second recommended processor core frequencies includes predicting processor workloads and determining application specific predicted P and C states core level processor frequencies to produce the further optimized power savings and performance.

[0131] |6| The method described in any of |3 | to |5 |, wherein the generating the second recommended processor core frequencies includes performing workload clustering to identify a pattern for the core processors of the one or more clusters at a core level.

[0132] [7] The method described in [6], wherein the performing workload clustering is based on information including one or more of a particular traffic pattern, throughput, UE Connection Lost according to a particular time of a day.

[0133] [8] An aspect of this description is directed to a system configured to obtain metrics from servers in a network, perform N iterations using the metrics to train a model of an artificial intelligence engine until an output of the model converges, generate first recommended processor core frequencies for the core processors of the servers using the trained model to provide optimal power savings and performance for the core processors of the servers, and provide the trained model and the first recommended processor core frequencies for the core processors to the servers.

[0134] [9] The system described in [8], wherein the first recommended processor core frequencies are based on an analysis of the metrics to identify information including one or more of workload signatures, network traffic patterns, time of day patters, or Operating System (OS) and Kernel statistics. [0135] [10] The system described in any of [8] to [9], further configured to receive model input and the first recommended processor core frequencies to create a reinforced model, receive Key Performance Indicators (KPIs) from pre-implementation of the first recommended processor core frequencies and KPIs based on implementation of the first recommended processor core frequencies, compare the KPIs from pre-implementation of the first recommended processor core frequencies and KPIs based on implementation of the first recommended processor core frequencies to identify a delta, based on the delta, perform M iterations until an output of the reinforced model converges to generate second recommended processor core frequencies to provide further optimized power savings and performance, and provide the second recommended processor core frequencies to the servers to provide the further optimized power savings and performance.

[0136] [11] The system described in [10], further configured to generate the second recommended processor core frequencies by segmenting one or more clusters of the servers using unsupervised learning and determining the second recommended processor core frequencies for the one or more clusters to produce the optimal power savings for each of the one or more clusters, wherein the determining the second recommended processor core frequencies for the one or more clusters includes determining optimized P states and C state for core processors of the one or more clusters of the servers and choosing a combination of the P states and C states to produce the optimal power savings and performance for the core processors of each of the one or more clusters of the servers.

[0137] [12] The system described in any of [10] to [1 1], further configured to generate the second recommended processor core frequencies by predicting processor workloads and determining application specific predicted P and C states core level processor frequencies to produce the further optimized power savings and performance.

[0138] [13] The system described in any of [10] to [12], further configured to generate the second recommended processor core frequencies by performing workload clustering to identify a pattern for the core processors of the one or more clusters at a core level.

[0139] [14] The system described in any of [13], further configured to perform the workload clustering based on information including one or more of a particular traffic pattern, throughput, UE Connection Lost according to a particular time of a day.

[0140] [15] An aspect of this description is directed to a non-transitory computer-readable media having computer-readable instructions stored thereon, which when executed operations are performed for obtaining metrics from servers in a network, performing N iterations using the metrics to train a model of an artificial intelligence engine until an output of the model converges, generating first recommended processor core frequencies for the core processors of the servers using the trained model to provide an optimal power savings and performance for the core processors of the servers, and providing the trained model and the first recommended processor core frequencies for the core processors to the servers.

[0141] [16] The non-transitory computer-readable media described in [15], wherein the generating the first recommended processor core frequencies are based on an analysis of the metrics to identify information including one or more of workload signatures, network traffic patterns, time of day patters, or Operating System (OS) and Kernel statistics.

[0142] [17] The non-transitory computer-readable media described in any of [15] to [16], wherein the operations further include receiving model input and the first recommended processor core frequencies to create a reinforced model, receiving Key Performance Indicators (KPIs) from pre-implementation of the first recommended processor core frequencies and KPIs based on implementation of the first recommended processor core frequencies, comparing the KPIs from pre-implementation of the recommended processor core frequencies and KPIs based on implementation of the first recommended processor core frequencies to identify a delta, based on the delta, performing M iterations until an output of the reinforced model converges to generate second recommended processor core frequencies to provide further optimized power savings and performance, and providing the second recommended processor core frequencies to the servers to provide the further optimized power savings and performance.

[0143] [18] The non-transitory computer-readable media described in [17], wherein the generating the second recommended processor core frequencies includes segmenting one or more clusters of the servers using unsupervised learning and determining the second recommended processor core frequencies for the one or more clusters to produce the optimal power savings and performance for each of the one or more clusters, wherein the determining the second recommended processor core frequencies for the one or more clusters includes determining optimized P states and C state for core processors of the one or more clusters of the servers and choosing a combination of the P states and C states to produce the optimal power savings and performance for the core processors of each of the one or more clusters of the servers. [0144] [19] The non-transitory computer-readable media described in any of [17] to [18], wherein the generating the second recommended CPU core frequencies includes predicting processor workloads and determining application specific predicted P and C states core level processor frequencies to produce the further optimized power savings.

[0145] [20] The non-transitory computer-readable media described in any of [17] to [19], wherein the generating the second recommended processor core frequencies includes performing workload clustering to identify a pattern for the core processors of the one or more clusters at a core level, and wherein the performing workload clustering is based on information including one or more of a particular traffic pattern, throughput, or UE Connection Lost according to a particular time of a day.

[0146] Separate instances of these programs can be executed on or distributed across any number of separate computer systems. Thus, although certain steps have been described as being performed by certain devices, software programs, processes, or entities, this need not be the case. A variety of alternative implementations will be understood by those having ordinary skill in the art.

[0147] Additionally, those having ordinary skill in the art readily recognize that the techniques described above can be utilized in a variety of devices, environments, and situations. Although the embodiments have been described in language specific to structural features or methodological acts, the subject matter defined in the appended claims is not necessarily limited to the specific features or acts described. Rather, the specific features and acts are disclosed as exemplary forms of implementing the claims.

Claims

CLAIMS WHAT IS CLAIMED IS:

1. A method, comprising: obtaining metrics from servers in a network; performing N iterations using the metrics to train a model of an artificial intelligence engine until an output of the model converges; generating first recommended processor core frequencies for the core processors of the servers using the trained model to provide an optimal power savings and performance for the core processors of the servers; and providing the trained model and the first recommended processor core frequencies for the core processors to the servers.

2. The method of claim 1 , wherein the generating the first recommended processor core frequencies are based on an analysis of the metrics to identify information including one or more of workload signatures, network traffic patterns, time of day patters, or Operating System (OS) and Kernel statistics.

3. The method of claim 1 further comprising: receiving model input and the first recommended processor core frequencies to create a reinforced model; receiving Key Performance Indicators (KPIs) from pre-implementation of the first recommended processor core frequencies and KPIs based on implementation of the first recommended processor core frequencies; comparing the KPIs from pre-implementation of the first recommended processor core frequencies and KPIs based on implementation of the first recommended processor core frequencies to identify a delta; based on the delta, performing M iterations until an output of the reinforced model converges to generate second recommended processor core frequencies to provide further optimized power savings and performance; and providing the second recommended processor core frequencies to the servers to provide the further optimized power savings and performance.

4. The method of claim 3, wherein the generating the second recommended processor core frequencies includes segmenting one or more clusters of the servers using unsupervised learning and determining the second recommended processor core frequencies for the one or more clusters to produce the optimal power savings and performance for each of the one or more clusters, wherein the determining the second recommended processor core frequencies for the one or more clusters includes determining optimized P states and C state for core processors of the one or more clusters of the servers and choosing a combination of the P states and C states to produce the optimal power savings and performance for the core processors of each of the one or more clusters of the servers.

5. The method of claim 3, wherein the generating the second recommended processor core frequencies includes predicting processor workloads and determining application specific predicted P and C states core level processor frequencies to produce the further optimized power savings and performance.

6. The method of claim 3, wherein the generating the second recommended processor core frequencies includes performing workload clustering to identify a pattern for the core processors of the one or more clusters at a core level.

7. The method of claim 6, wherein the performing workload clustering is based on information including one or more of a particular traffic pattern, throughput, or UE Connection Lost according to a particular time of a day.

8. A system for improving user experience in multi-frequency layer deployments, wherein the system is configured to: obtain metrics from servers in a network; perform N iterations using the metrics to train a model of an artificial intelligence engine until an output of the model converges; generate first recommended processor core frequencies for the core processors of the servers using the trained model to provide an optimal power savings and performance for the core processors of the servers; and provide the trained model and the first recommended processor core frequencies for the core processors of the servers.

9. The system of claim 8, wherein the first recommended processor core frequencies are based on an analysis of the metrics to identify information including one or more of workload signatures, network traffic patterns, time of day patters, or Operating System (OS) and Kernel statistics.

10. The system of claim 8, wherein the system is further configured to: receive model input and the first recommended processor core frequencies to create a reinforced model; receive Key Performance Indicators (KPIs) from pre-implementation of the first recommended processor core frequencies and KPIs based on implementation of the first recommended processor core frequencies; compare the KPIs from pre-implementation of the first recommended processor core frequencies and KPIs based on implementation of the first recommended processor core frequencies to identify a delta; based on the delta, performing M iterations until an output of the reinforced model converges to generate second recommended processor core frequencies to provide further optimized power savings and performance; and provide the second recommended processor core frequencies to the servers to provide the further optimized power savings and performance.

11. The system of claim 10, wherein the system is further configured to generate the second recommended processor core frequencies by segmenting one or more clusters of the servers using unsupervised learning and determining the second recommended processor core frequencies for the one or more clusters to produce the optimal power savings and performance for each of the one or more clusters, wherein the determining the second recommended processor core frequencies for the one or more clusters includes determining optimized P states and C state for core processors of the one or more clusters of the servers and choosing a combination of the P states and C states to produce the optimal power savings and performance for the core processors of each of the one or more clusters of the servers.

12. The system of claim 10, wherein the system is further configured to generate the second recommended processor core frequencies by predicting processor workloads and determining application specific predicted P and C states core level processor frequencies to produce the further optimized power savings and performance.

13. The system of claim 10, wherein the system is further configured to generate the second recommended processor core frequencies by performing workload clustering to identify a pattern for the core processors of the one or more clusters at a core level.

14. The system of claim 10, wherein the system is further configured to perform workload clustering based on information including one or more of a particular traffic pattern, throughput, or UE Connection Lost according to a particular time of a day.

15. A non-transitory computer-readable media having computer-readable instructions stored thereon, which when executed operations are performed for: obtaining metrics from servers in a network; performing N iterations using the metrics to train a model of an artificial intelligence engine until an output of the model converges; generating first recommended processor core frequencies for the core processors of the servers using the trained model to provide an optimal power savings and performance for the core processors of the servers; and providing the trained model and the first recommended processor core frequencies for the core processors to the servers.

16. The non-transitory computer-readable media of claim 15, wherein the generating the first recommended processor core frequencies are based on an analysis of the metrics to identify information including one or more of workload signatures, network traffic patterns, time of day patters, or Operating System (OS) and Kernel statistics.

17. The non-transitory computer-readable media of claim 15, wherein the operations further include: receiving model input and the first recommended processor core frequencies to create a reinforced model; receiving Key Performance Indicators (KPIs) from pre-implementation of the first recommended processor core frequencies and KPIs based on implementation of the first recommended processor core frequencies; comparing the KPIs from pre-implementation of the first recommended CPU core frequencies and KPIs based on implementation of the first recommended CPU core frequencies to identify a delta; based on the delta, performing M iterations until an output of the reinforced model converges to generate second recommended processor core frequencies to provide further optimized power savings and performance; and providing the second recommended processor core frequencies to the servers to provide the further optimized power savings and performance.

18. The non-transitory computer-readable media of claim 17, wherein the generating the second recommended processor core frequencies includes segmenting one or more clusters of the servers using unsupervised learning and determining the second recommended processor core frequencies for the one or more clusters to produce the optimal power savings and performance for each of the one or more clusters, wherein the determining the second recommended processor core frequencies for the one or more clusters includes determining optimized P states and C state for core processors of the one or more clusters of the servers and choosing a combination of the P states and C states to produce the optimal power savings and performance for the core processors of each of the one or more clusters of the servers.

19. The non-transitory computer-readable media of claim 17, wherein the generating the second recommended processor core frequencies includes predicting processor workloads and determining application specific predicted P and C states core level processor frequencies to produce the further optimized power savings and performance.

20. The non-transitory computer-readable media of claim 17, wherein the generating the second recommended processor core frequencies includes performing workload clustering to identify a pattern for the core processors of the one or more clusters at a core level, and wherein the performing the workload clustering is based on information including one or more of a particular traffic pattern, throughput, or UE Connection Lost according to a particular time of a day.