Internet service provider-based access layer system for peer-to-peer content delivery network

MPRD addresses HoL blocking and OOO packet delivery in PCDNs by optimizing multipath transport and deploying low-power cache nodes within ISP networks, improving throughput and reducing energy consumption.

WO2026236579A1PCT designated stage Publication Date: 2026-11-12PORTLAND STATE UNIV
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
PCT/US2026/027218
Authority / Receiving Office
WO · WO
Patent Type
Applications
Current Assignee / Owner
Priority Date
2025-05-08
Filing Date
2026-05-07
Publication Date
2026-11-12

AI Technical Summary

Technical Problem

Existing Peer-to-Peer Content Delivery Networks (PCDNs) face challenges such as Head-of-Line (HoL) blocking and Out-Of-Order (OOO) packet delivery, which degrade multipath throughput, and traditional content delivery models are energy-inefficient and unsustainable due to high energy consumption and network congestion.

Method used

A receiver-driven distributed transport protocol, MPRD, is introduced to optimize multipath transport by intelligently coordinating file segment downloads across multiple edge servers, combined with low-power cache nodes deployed within ISP networks to cache frequently accessed content, using Multipath Parallel Reverse Segment Download (MPRD) and intelligent caching strategies.

Benefits of technology

MPRD significantly improves client throughput, reduces CDN load, and lowers energy consumption by distributing content closer to users, enhancing video streaming quality and reducing network congestion.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure US2026027218_12112026_PF_FP_ABST
    Figure US2026027218_12112026_PF_FP_ABST
Patent Text Reader

Abstract

Embodiments herein relate to MPRD (Multipath Parallel Reverse Segment Download), a receiver-driven distributed transport protocol designed to overcome limitations in Peer-to-Peer Content Delivery Network (PCDN) architectures in caching and distributing content. MPRD is built on top of Transmission Control Protocol (TCP) and rethinks traditional multipath transport by using single-path TCP at each edge server. In MPRD, the client optimizes the starting point and direction of file segments to be downloaded by each single-path TCP connection in order to eliminate the inefficiencies of existing multipath schedulers.
Need to check novelty before this filing date? Find Prior Art

Description

Attorney Docket No.: 125512- 297358INTERNET SERVICE PROVIDER-BASED ACCESS LAYER SYSTEM FOR PEER-TO- PEER CONTENT DELIVERY NETWORK CROSS-REFERENCE TO RELATED APPLICATIONS

[0001] This application claims the benefit of U.S. provisional patent application no.63 / 802,473, filed May 8, 2025, entitled “Internet Service Provider-Based Access Layer System For Peer-To-Peer Content Delivery Network,” and incorporated herein by reference.STATEMENT REGARDING FEDERALLY SPONSORED RESEARCH OR DEVELOPMENT

[0002] This invention was made with government support under contract 1942305, awarded by the National Science Foundation. The government has certain rights in the invention.BACKGROUND

[0003] Internet content providers (e.g., video streaming, web services, file downloads) are increasingly deploying Peer-to-Peer Content Delivery Networks (PCDNs) to cache and distribute content by leveraging spare compute and storage resources at the network edge. This shift is primarily driven by the high costs associated with building and maintaining large-scale, centralized Content Delivery Network (CDN) architectures. A PCDN architecture can use multiple edge devices (e.g., access points, smart home gadgets) to distribute data in parallel to each client to increase the communication bandwidth and robustness to link / server failures. However, existing distributed multipath transport schedulers face significant challenges, such as Head-of-Line (HoL) blocking and Out-Of-Order (000) packet delivery, which can degrade multipath throughput to levels even lower than using a single path alone.BRIEF DESCRIPTION OF THE DRAWINGS

[0004] The embodiments of the disclosure will be understood more fully from the detailed description given below and from the accompanying drawings of various embodiments of the disclosure, which, however, should not be taken to limit the disclosure to the specific embodiments, but are for explanation and understanding only.

[0005] FIG. 1 depicts an example network architecture 100 of a fixed broadband Internet Service Provider (ISP), in accordance with various embodiments.Attorney Docket No.: 125512- 297358

[0006] FIG. 2 depicts an example network architecture 200 of a broadband ISP with cache clusters deployed in incentivized homes (darkened homes), in accordance with various embodiments.

[0007] FIG. 3 depicts an example PCDN 300 consistent with FIG. 1, where Head-of-Line (HoL) blocking and Out-Of-Order (OOO) packets can occur, in accordance with various embodiments.

[0008] FIG. 4 depicts an example PCDN 400 consistent with FIG. 2, where different pieces of content come from different servers and paths, avoiding HoL blocking and OOO packets, in accordance with various embodiments.

[0009] FIG. 5 depicts an example architecture 500 of a fixed broadband ISP integrated with cache clusters and an access-layer orchestrator, in accordance with various embodiments.

[0010] FIG. 6 depicts an example computing device 600 representing any of the nodes or components described herein, in accordance with various embodiments.

[0011] FIG. 7 depicts an example of cache cluster software 700, in accordance with various embodiments.

[0012] FIG. 8 depicts an example network architecture 800 consistent with FIG. 2 with cache clusters deployed in incentivized homes, where a user in a target home 130a receives content such as a video from incentivized homes 130b- 133b, in accordance with various embodiments.

[0013] FIG. 9 depicts an example satellite communication architecture 900, with multiple cache nodes transmitting data such as video segments to a SAT user, in accordance with various embodiments.

[0014] FIG. 10 depicts a system architecture 1000 for cache clusters and access layer orchestrator for a mobile network such as a 5G or Long-Term Evolution (LTE) network, in accordance with various embodiments.

[0015] FIG. 11 depicts how a CDN can use a hierarchical structure 1100 with an origin server 1110 at the root, regional data centers 1120 in the middle, and edge servers 1130 close to the end users / clients 1140, in accordance with various embodiments.

[0016] FIG. 12 depicts how a PCDN can employ a decentralized architecture 1200, operating standalone or in hybrid mode to reduce CDN load, in accordance with various embodiments.

[0017] FIG. 13 depicts an architecture 1300 in which a tracker 1310 lists edge servers and their stored content, guiding clients (edge devices 1330) to the appropriate servers (PCDN users 1320) for downloads, in accordance with various embodiments.Attorney Docket No.: 125512- 297358

[0018] FIG. 14 depicts a bar chart showing the throughput of each server to the client, both in isolation and when serving the client jointly, in accordance with various embodiments.

[0019] FIG. 15 depicts a plot of Y-channel Peak Signal-to-Noise Ratio (Y-PSNR), a visual quality metric, versus encoding data rate, for an example video entitled “Big Buck Bunny,” in accordance with various embodiments.

[0020] FIG. 16 depicts equations to compute the transmission window offset for each edge server, in accordance with various embodiments.

[0021] FIG. 17A depicts an initial placement of forward and reverse pointers at the client reception window, where the dashed parts represent the segments transmitted by the associated server, in accordance with various embodiments.

[0022] FIG. 17B depicts how, when forward and reverse pointer 1 cross over (i.e., finish transmitting the segments between them), MPRD reassigns them to help with transmitting segments in T Woffset2, and the blocks 1710 and 1720 show the segments received by the client, in accordance with various embodiments.

[0023] FIG. 17C depicts how, as reverse pointer 2 finishes transmitting its segments, MPRD reassigns it to be in between the forward pointer and reverse pointer 1, in accordance with various embodiments.

[0024] FIG. 18 depicts client and server processes on two separate machines 1810 and 1820, in accordance with various embodiments.

[0025] FIG. 19 depicts how Linux TC can control bandwidth, delay, and loss for each path, in accordance with various embodiments.

[0026] FIG. 20A depicts a bar chart of throughput across different schemes as a function of number of edge servers in a low throughput configuration, where no server experiences packet loss, in accordance with various embodiments.

[0027] FIG. 20B depicts a bar chart of throughput results when edge server one experience 10% packet loss rate, in accordance with various embodiments.

[0028] FIG. 20C depicts a bar chart of CDN load as a function of number of servers (CDN server + zero to four edge servers), where no server experiences packet loss, in accordance with various embodiments.

[0029] FIG. 20D depicts a bar chart of throughput across different schemes in a high throughput configuration, where no server experiences packet loss, in accordance with various embodiments.

[0030] FIG. 20E depicts a bar chart of throughput results when edge server one experiences 10% packet loss rate, in accordance with various embodiments.Attorney Docket No.: 125512- 297358

[0031] FIG. 20F depicts a bar chart of CDN load as a function of number of servers (CDN server + zero to four edge servers), where edge server one experience 10% packet loss rate, in accordance with various embodiments.

[0032] FIG. 21 A depicts a bar chart of the percentage of pauses for different data rates of Big Buck Bunny video, in accordance with various embodiments.

[0033] FIG. 2 IB depicts a bar chart of PSNR for different data rates of Big Buck Bunny video, in accordance with various embodiments.

[0034] FIG. 21C depicts a bar chart of file download time for different data rates of Big Buck Bunny video, in accordance with various embodiments.

[0035] FIG. 22 depicts an example communication architecture 2200 including a micro point of presence (M-PoP) 2210 in communication with an optical line termination (OLT) device 2220, in accordance with various embodiments.

[0036] FIG. 23 depicts an example communication architecture 2300 including the M-PoP 2210 at a central office 2310, in accordance with various embodiments.

[0037] FIG. 24 depicts an example system architecture of the M-PoP 2210, in accordance with various embodiments.

[0038] FIG. 25 depicts an example system architecture 2500 with M-PoP modules / nodes using Kubemetes clusters, in accordance with various embodiments.

[0039] FIG. 26 depicts an example micro CDN architecture 2600 using Kubernetes pods and clusters, in accordance with various embodiments.

[0040] FIG. 27 depicts an example communication architecture 2700 in which the M-PoP 2210 is connected directly to an Ethernet port, in accordance with various embodiments.

[0041] FIG. 28 depicts an example communication architecture 2800 in which the M-PoP 2210 is connected to an Ethernet port via a Virtual Local Area Network (VLAN) 2810, in accordance with various embodiments.DETAILED DESCRIPTION

[0042] As mentioned at the outset, various challenges are encountered in caching and distributing content in computer networks such as the internet.

[0043] The internet is growing at a massive pace, driven by several powerful trends -Cloud Computing, OTT (Over-the-Top) video streaming, the explosion of mobile usage, Gaming, Artificial Intelligence (Al), and Edge Computing. These technologies are changing how creation, sharing, and consumption of digital content. As more users stream videos, use cloud services, and rely on mobile applications, the demand for faster and more scalableAttorney Docket No.: 125512- 297358infrastructure has skyrocketed. However, the current content delivery model is increasingly unsustainable due to its high energy consumption, limited scalability, and inefficiencies in handling redundant content delivery.

[0044] To support this demand, the world has turned to hyperscale data centers - huge facilities run by services such as Amazon Web Services (AWS), Google Cloud, and Microsoft Azure. These data centers can store and process petabytes of data, serve millions of users at once, and grow or shrink based on real-time demand. They form the backbone of the modern digital economy. However, there's a major downside to hyperscale data centers: they consume an enormous amount of power. These facilities need electricity not only to run thousands of servers but also to power cooling systems that prevent overheating and to ensure that everything runs around the clock (24 / 7). This has made them one of the biggest contributors to global energy usage. As more digital services are used daily, the environmental impact of these data centers continues to grow, raising concerns about sustainability, carbon emissions, and the need for greener solutions.

[0045] While expanding hyperscale infrastructure can’t be entirely avoided, not all internet traffic needs such powerful computing resources. A lot of online content - especially videos - is watched repeatedly by many users and doesn’t need to be delivered directly from the main data center every time. Instead, it can be cached closer to where users are located. This means storing popular content temporarily on nearby servers to reduce energy use, network traffic, and cost.

[0046] In fact, 60-80% of all consumer internet traffic is video streaming. In local neighborhoods, studies show that 20-40% of videos watched are the same, and for highly trending content like CNN or ESPN clips, this number goes up to 50%. That means a lot of video content is being repeatedly requested by nearby users.

[0047] This is where a Peer-to-Peer CDN (PCDN) comes into the picture. PCDN is a model where devices that are already playing a video (or some other content) store that content temporarily and then share it with nearby devices that are watching the same video. This helps reduce the load on central servers and saves bandwidth and energy.

[0048] However, direct sharing between user devices raises several security and policy concerns:• Is the content authentic?• Could malware be spread?• What about Digital Rights Management (DRM) and copyright?Attorney Docket No.: 125512- 297358

[0049] Because of these issues, most content providers and ISPs do not support this type of peer sharing. Also, ISPs often have strict security filters (called ingress / egress filters) that block or limit such direct connections between users.

[0050] So instead of true peer-to-peer (P2P) sharing, most big video platforms like Netflix, YouTube, and Disney-i- use a different approach. They deploy edge nodes — servers placed inside ISP networks at strategic points called Points of Presence (PoPs) or in MultiAccess Edge Computing (MEC) locations as shown in FIG. 1. These edge nodes store frequently watched content locally, so it doesn’t have to travel all the way from the main data center.

[0051] For example, when you watch a video at home, your request goes from your home network to your local aggregation point, and then to the nearest point of presence (PoP) or MEC location where the content might already be cached. If it’ s there, the video is served from that point, which saves on upstream bandwidth (also known as uplink traffic). These edge caches are often called peering nodes.

[0052] While this helps reduce traffic back to the main data center and lowers bandwidth costs, these peering nodes still operate like small data centers, consuming significant energy and infrastructure. So while they help with bandwidth savings, they do not fully solve problems like:• Network congestion• ISP bandwidth expansion• High energy use

[0053] Herein, the key challenges faced by the traditional PCDN model are addressed. A new solution is proposed: deploying lightweight, energy-efficient edge nodes deeper inside the ISP network, closer to the end users. This approach is only possible through strong partnerships between ISPs and content providers. Together, they can build a smarter, more efficient content delivery network that reduces energy use, lowers cost, improves performance, and supports sustainable growth in digital traffic.

[0054] In an example implementation, MPRD (Multipath Parallel Reverse Segment Download) is introduced, a novel receiver-driven distributed transport protocol designed to overcome these limitations in PCDN architectures. MPRD is built on top of Transmission Control Protocol, (TCP) and rethinks traditional multipath transport by using single-path TCP at each edge server. In MPRD, the client optimizes the starting point and direction of file segments to be downloaded by each single-path TCP connection in order to eliminate the inefficiencies of existing multipath schedulers. A prototype of MPRD was implemented in theAttorney Docket No.: 125512- 297358Linux user space and demonstrates that it significantly improves client throughput and reduces the CDN load (when the CDN is augmented by PCDN edge servers). MPRD’s performance was also evaluated in improving video delivery quality and shows that it can substantially reduce the stalls and increase the PSNR visual quality by up to 35 dB.

[0055] The solutions provided herein address the shortcoming of other potential solutions.

[0056] For example, with Multipath Transmission Evaluation, several studies have analyzed Multipath TCP (MPTCP) performance (e.g., in terms of throughput or energy efficiency), but their evaluations are largely confined to an architecture where a single server communicates with a single client over multiple paths. This setup differs from the PCDN architectures explored in this paper. Other research has proposed new schedulers built on MPTCP for PCDN architectures. The performance of MPRD was evaluated against the state-of-the-art in this domain (P-Scheduler) and demonstrated that MPRD delivers superior performance across various metrics, including total throughput and PSNR, and for different numbers of edge servers. Furthermore, MPRD is implemented in the Linux user space.

[0057] With Multipath Transmission Schedulers, the Linux kernel includes several MPTCP schedulers by default, such as RoundRobin, Minimum Round-Trip Time (minRTT), and BLEST. RoundRobin is well-suited for scenarios where all paths have similar characteristics (e.g., delay and data rate), while minRTT and BLEST are designed to perform effectively across more heterogeneous links. Additionally, the research community has proposed various schedulers, such as those addressing challenges posed by heterogeneous paths, schedulers leveraging differences in subflow RTTs, and others aimed at improving MPTCP performance for specific use cases. However, all these solutions have been designed and evaluated for single- server- to- single-client multipath architectures. MPRD’s performance was compared against minRTT (a default Linux MPTCP scheduler) and P-Scheduler, which represents the state-of-the-art for PCDN architectures. While experiments were also conducted with the BLEST scheduler, the results were omitted as BLEST consistently achieved performance levels between minRTT and P-Scheduler.

[0058] Another potential solution involves Content Delivery Architectures. CDNs such as Akamai and Cloudflare are systems of servers that deliver web content and media to users based on their geographic location, ensuring low latency and high availability. PCDNs extend this concept by leveraging end-user devices as part of the delivery infrastructure, enabling cost-effective scalability and better utilization of edge resources. Hybrid CDNs combine the strengths of traditional CDNs and PCDNs, using centralized servers for reliable delivery andAttorney Docket No.: 125512- 297358peer-to-peer networks for efficient resource utilization and scalability, particularly during traffic surges or in regions with limited server presence. MPRD is a transport protocol designed for multipath architectures, with a focus on increasing throughput in PCDNs or reducing CDN load in hybrid architectures.

[0059] These and other features will be further apparent in view of the following discussion.

[0060] FIG. 1 depicts an example network architecture 100 of a fixed broadband ISP, in accordance with various embodiments. The architecture includes a set of homes 110 including subsets 130-133 and 140-143. Each subset includes a plurality of homes such as 64 homes. For example, the subset 130 includes an example home 130a. The subsets communicate with an aggregation layer node 160, that can include aggregation switches and a central unit (CU), via splitters. For examples, the subsets 130-133 communicate via local splitters 111-114, respectively, and via a shared splitter 115 and communication path 116. The subsets 140-143 communicate via local splitters 121-124, respectively, and via a shared splitter 125 and communication path 126.

[0061] The aggregation layer nodes communicates with a node 170, that can include an ISP POP, an edge node or a gateway multi access computing (MEC) node. The node 170 can receive data from a CDN or cache boxes. The node 170 communicates with a node 180 that can include an upstream provider or 5G core. The node 180 communicates with the internet 190.

[0062] FIG. 2 depicts an example network architecture 200 of broadband ISP with cache clusters deployed in incentivized homes (darkened homes), in accordance with various embodiments. In this example, there is one incentivized home in each subset. However, other approaches are possible. For example, there could be more than one incentivized home in a subset, or some subsets may have not incentivized homes. Subsets 130-133 include incentivized homes 130b-133b, respectively, and subsets 140-143 include incentivized homes 140b- 143b, respectively.

[0063] FIG. 3 depicts an example PCDN 300 consistent with FIG. 1, where HoE blocking and 000 packets can occur, in accordance with various embodiments. The PCDN includes a client 310, edge serverl (320), edge server 3 (330) and edge server 3 (340). The client and the edge servers can each store up to eight data segments in respective storage locations. For example, the client 310 includes first through sixth storage locations 311-316, respectively, the edge server 1 includes first through sixth storage locations 321-326, respectively, the edge server 2 includes first through sixth storage locations 331-336, and theAttorney Docket No.: 125512- 297358edge server 3 includes first through sixth storage locations 341-346. The packets are received in a forward or left-to-right order at the client.

[0064] Arrows 347, 337 and 327 represent a data transfer from the storage location 341 to the storage location 315, from the storage location 331 to the storage location 313, and from the storage location 321 to the storage location 311, respectively.

[0065] FIG. 4 depicts a PCDN where different pieces of content come from different servers and paths, where HoL blocking and 000 packets are avoided, in accordance with various embodiments. The PCDN includes the client 310, edge serverl (320), and edge server 2 (330) of FIG. 3, and a CDN server 440. The CDN server can also store up to eight data segments in respective storage locations. For example, the CDN server includes first through sixth storage locations 441-446, respectively.

[0066] Arrows 447, 448 and 449 represent a data transfer from the storage location 441 to the storage location 316, from the storage location 324 to the storage location 313, and from the storage location 336 to the storage location 311, respectively. The packets are received in a reverse or right-to-left order at the client.

[0067] FIG. 5 depicts an example architecture 500 of a fixed broadband ISP integrated with cache clusters and an access-layer orchestrator, in accordance with various embodiments. A top layer of the architecture includes an aggregation layer 505. A next layer of the architecture includes an access layer orchestrator 510. A next layer of the architecture includes splitters 515. A final layer of the architecture includes nodes 520, 521, ..., 522 representing ONT (an optical termination network), a cable modem or a Digital Subscriber Line (DSL) modem, in a number N of different homes denoted as Homel, Home2, .., HomesN, respectively. The final layer further includes nodes 530, ..., 531 representing a number M of cache clusters denoted as Cashe clusters 1, ..., Cashe clusters 1, respectively, in incentivized homes denoted as incentivized homel, ..., incentivized homeM.

[0068] FIG. 6 depicts an example computing device 600 representing any of the components described herein, in accordance with various embodiments. The computing device includes a tangible, non-transitory memory 605 (computer-readable medium) that is configured to store instructions for execution by a processor 610 to perform the functions described herein. The computing device may further include a communications interface 615 to communicate with other computing devices. The components can be implemented on one or more circuit boards, for example. Components on user premises may be housed in a console.

[0069] FIG. 7 depicts an example of cache cluster software 700, in accordance with various embodiments. The software includes containers 701, 702, 703, ..., 704. EachAttorney Docket No.: 125512- 297358container can include a CDN, Software as a Service (SaaS), Infrastructure as a Service (laaS), and / or Platform as a Service (PaaS).

[0070] FIG. 8 depicts an example network architecture 800 consistent with FIG. 2 with cache clusters deployed in incentivized homes, where a user in a target home 130a receives content such as a video from incentivized homes 130b- 133b, in accordance with various embodiments.

[0071] Referring to the target home 130a with a slant line pattern, aggregation occurs using MPRD at the client. Four incentivized homes in this example provide video chunks from their cache clusters to the target home. Additionally, video chunks are provided to the target home from the CDN / edge node (node 170). In an example implementation, about four users per home can be supported since there are four incentivized homes providing video or other content.

[0072] In this example, one incentivized home in each of the subsets 131-133 are used to transmit content to the target home in the subset 130. In particular, incentivized home 131b, 132b and 133b are used to transmit content to the target home 130a. In another example, an incentivized home is in the same subset as the target home. In another example, multiple incentivized homes is a subset are used to provide content to a target home.

[0073] The solutions herein introduce a secure, energy-efficient, and scalable access layer PCDN architecture that addresses the limitations of today’s internet infrastructure. With the exponential growth in digital traffic — driven by video streaming, cloud computing, enterprise SaaS applications, mobile access, and real-time collaboration — traditional content delivery models built around centralized hyperscale data centers are reaching a breaking point. These systems consume vast amounts of energy, are prone to redundancy in data transmission, and often struggle with network congestion and latency due to the physical distance between users and content sources.

[0074] One of the most pressing challenges in deploying PCDN has been security and authentication. Traditional peer-to-peer models allow user devices to cache and serve content to one another, but they introduce serious risks — such as the potential for spreading malware, violating content ownership rights, and breaching digital rights management (DRM) policies. The question then becomes: how can content be cached and delivered closer to the user while maintaining full control, security, and trust?

[0075] The solutions herein answer that question by shifting the PCDN paradigm. Instead of relying on unmanaged end-user devices, it proposes a distributed caching system built in partnership with ISPs and content providers. Low-cost, low-power cache nodes —Attorney Docket No.: 125512- 297358consuming less than 15-30 W, for example (can be higher power-consuming processors and higher cost as well) and powered by processors such as Intel N150, RK3588 (Fuzhou Rockchip Electronics Co., Ltd.) or similar processors — are deployed at the access layer of ISP networks, specifically within select residential units called incentivized homes. These devices, known as cache clusters, mimic typical home devices but are fully controlled and authenticated by the ISP or content provider. This eliminates the security and trust barriers that have long hindered peer-based content sharing.

[0076] These cache clusters form a distributed, fault- tolerant mesh of local storage and delivery points. Each node is equipped with Random Access Memory (RAM) or Synchronous Dynamic RAM (SDRAM), Non-Volatile Memory Express (NVMe), Solid-State Drive (SSD) storage, and Ethernet connectivity. Collectively, they cache and serve frequently accessed content — such as trending videos, portal data, Voice over Internet Protocol (VoIP) packets, and application payloads — directly to nearby users. This dramatically reduces uplink bandwidth demand, alleviates congestion in the ISP network, lightens the load on CDN edge nodes and centralized servers, and significantly reduces overall energy consumption. The system is redundant and self-healing, ensuring high availability and continuous operation across residential environments.

[0077] Video Streaming. One of the most powerful applications of this architecture is in video streaming, that constitutes the majority of today’s internet traffic. By caching and delivering the most viewed video content from multiple coordinated cache nodes, the system can support higher-quality streaming — including High-definition (HD) and even 4K video — without requiring ISPs to increase backbone bandwidth or content providers to scale up server capacity. Even with the existing ISP infrastructure, distributing streaming workloads across local cache nodes significantly reduces the load on central servers in data centers, leading to measurable energy savings and lower carbon emissions. By shifting the bulk of video delivery away from centralized hyperscale facilities and into the edge of the network, the system not only improves efficiency but also supports broader goals of sustainability and green computing.

[0078] For instance, consider a typical ISP network capable of streaming HD (1080p) video at around 8 Mbps. Under conventional infrastructure, this limits the quality and scalability of content delivery. However, when video segments are distributed across multiple cache nodes in the access network, the system can aggregate their throughput. For example, with four cache nodes and a CDN server each capable of streaming at 8 Mbps, the network can collectively deliver video at 35-45 Mbps, that is sufficient for smooth 4K streaming — without any upgrades to ISP infrastructure or backend servers.Attorney Docket No.: 125512- 297358

[0079] To achieve this level of throughput aggregation, the solutions herein introduce an intelligent receiver-driven coordination mechanism called the Multi-Path Reverse and Forward Direction (MPRD) scheduler. See further details in IV. A. Overall Architecture.

[0080] Beyond video streaming. A growing share of traffic comes from cloud-based enterprise applications, including Software as a Service (SaaS) platforms such as Customer Relationship Management (CRM) platforms (e.g., Salesforce, Zendesk), eCommerce services, cloud productivity suites (e.g., Office 365, Google Workspace, Zoho Workplace), Al inference workloads, augmented and virtual reality (AR / VR) applications, gaming, and software updates — all of which demand high availability and low latency.

[0081] Additionally, businesses increasingly rely on Infrastructure as a Service (laaS) and Platform as a Service (PaaS) offerings to deploy, manage, and scale their own applications. These services often involve frequent access to cloud-hosted virtual machines, databases, and APIs, that generate both predictable and dynamic traffic flows. Efficient access to these services, especially for remote workers and distributed teams — requires stable, responsive, and high-throughput network connections.

[0082] Internal communications within organizations — such as VoIP calls, video conferencing, and team chat applications — typically remain within the ISP’s infrastructure but still contribute significantly to network load. Furthermore, credential-based services such as email platforms, social networking sites, and document management systems generate a blend of dynamic and static content. Elements like profile pictures, attachments, and dashboards are repeatedly accessed and can benefit greatly from local caching.

[0083] In each of these use cases, localized edge caching within the ISP network offers a compelling advantage. By serving frequently requested or redundant content directly from nearby, ISP-managed cache nodes (hosted at Incentivized homes, or outdoor cabinets, near / along with the headend), the system dramatically reduces backhaul strain, lowers latency, and improves overall user experience while conserving energy and infrastructure costs at the core.

[0084] The deployment of but not limited to low-cost, low-power processors as cache nodes at the access layer is proposed, in collaboration with ISPs — including fixed broadband, mobile, and satellite — and content providers such as YouTube, Netflix, and Amazon Prime etc.,. These cache nodes aim to reduce the load on traditional CDN and edge servers in data centers, thereby lowering energy consumption and minimizing carbon footprints. Content such as videos, software updates, and large files will be cached across multiple nodes and streamedAttorney Docket No.: 125512- 297358to clients using the Multipath Reverse Segment Download (MPRD) algorithm, which improves throughput and efficiency by sourcing segments from multiple cache points simultaneously.

[0085] These cache nodes can be extended to host micro CDNs for a wide range of services and use cases, including ecommerce, gaming, education, Software as a Service (SaaS), Infrastructure as a Service (laaS), Platform as a Service (PaaS), media and entertainment, software downloads and updates, public, government and private sector websites, developer platforms like GitHub and GitLab, and core network services like Domain Name System (DNS), Dynamic Host Configuration Protocol (DHCP), and database caching. Micro CDNs, typically hosted using but not limited to Nginx, serve as lightweight, distributed edge servers that can significantly improve performance and resilience.

[0086] The remainder of this section is organized as follows:

[0087] 1. Cache clusters for fixed broadband ISPs

[0088] 2. Cache clusters for satellite ISPs

[0089] 3. Cache clusters for mobile ISPs

[0090] 4. MPRD algorithm for segment downloads from multiple nodes or cache clusters

[0091] Cache Clusters for Fixed ISP Broadband Deployment

[0092] Background. Fixed ISPs-which offer broadband services — typically use one of three main technologies to deliver the internet to homes and businesses: Cable, Optical Fiber, or Digital Subscriber Line (DSL). In a typical fixed broadband network, the architecture follows a layered structure that helps ISPs efficiently manage and scale internet access for many users.

[0093] At the customer’s home or business, there is usually a device known as Customer Premises Equipment (CPE) — such as a DSL modem, cable modem, or Optical Network Terminal (ONT) — depending on the type of connection. This device links the user's computers, phones, or Wi-Fi routers to the ISP's network. These CPE devices connect to the access layer through splitters, which divide and manage the network signal. A single splitter typically serves 32 or 64 homes — known as a 32: 1 or 64: 1 split ratio. This marks the first stage where traffic from multiple homes is combined.

[0094] From there, traffic flows to the aggregation layer, where multiple splitters feed into centralized network nodes. These nodes may include:

[0095] • Optical Line Terminals (OLT) for fiber networks,

[0096] • Cable Modem Termination Systems (CMTS) for cable networks, andAttorney Docket No.: 125512- 297358

[0097] • Broadband Remote Access Servers (BRAS) or Broadband Network Gateways (BNG) for DSL and general broadband services.

[0098] These aggregation devices collect and manage user traffic and pass it further into the network. Next, data moves into the core network, traveling from the aggregation layer to the ISP's Point of Presence (POP) or directly to an Internet Exchange Point (IXP). These are key hubs where ISPs connect to upstream providers or exchange traffic with other networks, enabling access to the broader internet.

[0099] The typical data flow from a user's device to the internet follows this path: Home Modem (CPE) «-> Splitter (Access Layer) «-> Aggregation Device (OLT / CMTS / BRAS) POP Internet Exchange / Upstream. This structured design ensures reliable and scalable internet access across neighborhoods, cities, and regions.

[0100] An architecture is proposed where but not limited to low-cost, low-power computing units — such as Intel N150 / N100, RK3588 or similar — are strategically deployed between the splitter and aggregation layers, as well as at either customer premises, outdoor cabinets or near / alone with headends (such as Optical Line Termination etc.). These devices play two roles: as access-layer orchestrators in the network and as cache clusters either within customer homes, also known as incentivized homes, outdoor cabinets or near / along headend. The orchestrators manage and deploy workloads dynamically across the network using orchestration tools like Kubernetes, while the cache clusters host lightweight containers running diverse applications based on local or regional traffic demands. This distributed architecture supports various services such as video streaming, web hosting, CRM, and more, optimizing network efficiency and bringing content closer to the end user.

[0101] In an example implementation:

[0102] 1. Low-cost, low-power computing units like Intel N150 / N100 and RK3588 are deployed between the splitters and aggregation layers or near / integrated with headend, referred to as access-layer orchestrators.

[0103] 2. Similar computing units are also deployed at the customer premises, where they function as cache clusters, shown in FIGs. 5-7. Homes hosting these units are called incentivized homes.

[0104] 3. Cache clusters typically run lightweight operating systems such as Linux or Windows, and host lightweight containers.

[0105] 4. Access-layer orchestrators also run operating systems like Linux or Windows, and include orchestration platforms such as Kubernetes, which dynamically deploy workloads to the cache clusters.Attorney Docket No.: 125512- 297358

[0106] 5. These orchestrators include load balancers such as but not limited to Nginx, Traefik, or HAProxy.

[0107] 6. Cache clusters / orchestrator also serve as DNS servers, using tools like DNSmasq.

[0108] 7. Cache clusters in either incentivized homes, outdoors cabinets or near / along with headend act as worker nodes to the access-layer orchestrators.

[0109] 8. The workloads on cache clusters can include any type of application, such as video streaming, CRM platforms (e.g., Salesforce, Zendesk), eCommerce services, cloud productivity suites (e.g., Office 365, Google Workspace, Zoho Workplace), Al inference workloads, augmented and virtual reality (AR / VR) applications, gaming, and software updates — all of which demand high availability and low latency.

[0110] 9. Workloads are deployed on cache clusters based on resident traffic patterns, and may also be coordinated with neighboring orchestrators, that can belong to the same or different ISPs.

[0111] 10. Each container running on the cache clusters serves a specific workload, such as but not limited to ABC News, CNN, YouTube, Netflix, Salesforce, etc.

[0112] Video Streaming with MPRD for Fixed Broadband ISP(s):

[0113] When a user requests video content, the system uses a workload orchestrator to route the request to appropriate micro CDN containers based on the type of content provider (e.g., YouTube, Netflix). Frequently accessed videos are cached in local cache clusters using intelligent caching strategies like Feast Recently Used (ERU), artificial intelligence (AI) / machine learning (ME)-based or similar algorithms. Each content provider has dedicated workload containers or / and shared storage with clear boundaries. If the requested video is already cached (cache hit), it's served directly from the local cache. Otherwise (cache miss), the video is fetched from external CDN or edge nodes, stored in the cache, and then delivered to the user. Videos can be delivered from one or multiple cache clusters. When served from multiple clusters, the user device employs a receiver-driven MPRD (Multipath Parallel Receiver-Driven) algorithm to download segments from multiple sources efficiently.

[0114] In an example implementation:

[0115] 1. Users requesting video content will send their request to the workload orchestrator, which then forwards it to the container running the appropriate micro CDN server based on but not limited load and other factors decided by the Kubernetes algorithm. For example, if a user requests a YouTube video, the workload orchestrator intercepts the request and assigns it to the work nodes hosting the YouTube workload containers.Attorney Docket No.: 125512- 297358

[0116] 2. Frequently viewed videos are stored in cache clusters. The micro CDN uses caching algorithms such as Least Recently Used (LRU), AFML-based algorithms or similar to manage and optimize video caching.

[0117] 3. Videos from each content provider — such as YouTube, Netflix, Disney+, etc. — are stored either in their respective workload containers or in shared storage with proper separation. These videos may reside in cache clusters or in centralized storage.

[0118] 4. On a cache hit, videos stored locally are delivered directly to the user. On a cache miss, the content is fetched from the CDN or edge nodes, then cached locally and delivered to the user.

[0119] 5. User-requested videos can be served from a single cache cluster or multiple cache clusters, depending on availability and system optimization.

[0120] 6. When a user is served from multiple cache clusters, the client uses but not limited to a receiver-driven MPRD (Multipath Parallel Receiver-Driven) algorithm to download different segments from those clusters. For example, if a user from the target home 130a in FIG. 8 requests a video, they may be served by four incentivized homes 130b- 133b. The user at the target home uses the MPRD algorithm to receive all video segments, as described in the MPRD section.

[0121] Cache Clusters for Satellite Internet Service Provider Deployment

[0122] Background - Satellite Communication Architecture and Deployment Challenges. Satellite communication systems have become important to bridging the digital divide in rural and remote regions, especially where terrestrial fiber infrastructure is economically unfeasible. These systems use geostationary (GEO), medium earth orbit (MEO), and increasingly, low earth orbit (LEO) satellites to provide broadband internet, voice services, and Intemet-of-Things (loT) connectivity. LEO satellite constellations are particularly attractive due to their lower latency and higher throughput capabilities.

[0123] In a typical architecture, user devices such as home terminals or mobile VSATs communicate directly with LEO satellites via radio frequency uplinks. These satellites relay the signals to terrestrial ground stations (gateways), which are connected to internet cloud data centers. The data is processed, routed, or fetched from the internet, then sent back through the reverse path — from ground station to satellite, and finally back to the end user. This multihop communication flow can be summarized as shown in FIG. 9.

[0124] FIG. 9 depicts an example satellite communication architecture 900, with multiple cache nodes transmitting data such as video segments to a SAT user, in accordance with various embodiments. The architecture includes a satellite 910 in orbit and a number ofAttorney Docket No.: 125512- 297358ground stations of respective homes, including incentivized homes and non-incentivized homes. For example, the incentivized homes include homes 932a, 933a and 934a including satellite modems 932b, 933b and 934b coupled to satellite dishes 932, 933 and 934, respectively, and being associated with caches 932c, 933c and 934c, respectively. The nonincentivized homes include home 930a including satellite modem 930b coupled to satellite dish 930. Each satellite modem may have the ability to transmit data to and receive data from the satellite 910.

[0125] The architecture 900 further includes a satellite dish 937 coupled to a modem 937c of a utility shed 937b, and to a cloud data center 945.

[0126] The target home 930a is coupled to the respective satellite dish 930 and includes a satellite modem 930b. The satellite dishes 932, 933, 934 and 937 provide uploads via paths 932d, 933d, 934d and 937d, respectively. The target home receives data from each of the incentivized homes via a download stream 930c from the satellite 910:User Satellite Ground Station Cloud Data Center Ground Station Satellite User

[0127] Modem LEO satellites use advanced spot beam technology, dividing their coverage areas into multiple focused beams. A single satellite may support 20 to 60 spot beams, with 1,000 to 5,000 users per beam, depending on service density and bandwidth allocation. On the ground, each gateway station may support 50 to 100+ beams, enabling large-scale user connectivity across wide geographic footprints.

[0128] These systems operate across multiple frequency bands, depending on use case: 1. L-band and S-band for narrowband communication and resilient low-data-rate services (e.g., GPS, loT, satellite phones).2. Ku-band and Ka-band for high-speed internet access and broadband backhaul, especially in fixed home and enterprise deployments.3. Advanced constellations also use Orthogonal Frequency Division Multiplexing (OFDM) for its efficiency and compatibility with 5G and broadband systems.

[0129] However, this architecture introduces key performance and economic challenges:• High latency and limited throughput due to long round-trip paths between user, satellite, and centralized cloud infrastructure.• Network congestion and backhaul overload, especially during peak usage hours, as all data flows through central ground stations.Attorney Docket No.: 125512- 297358• Scalability limitations, where serving tens of thousands of users per satellite strains spectrum, processing, and ground infrastructure.

[0130] To address these bottlenecks, emerging solutions explore the use of edge nodes and cache boxes placed closer to the user — at the gateway or potentially even onboard the satellite — to reduce round-trip delays and optimize bandwidth usage. By offloading repeated content delivery and localized routing from the central data center, these distributed systems aim to provide lower latency, improve QoS (Quality of Service), and reduce operating costs.

[0131] The following architecture is proposed:

[0132] A satellite-based distributed caching and orchestration architecture uses low-cost, low-power computing units — such as but not limited to Intel N150 / N100, RK3588 or similar — deployed in incentivized homes equipped with satellite modems and antennas. These units serve dual roles: some function as access-layer orchestrators and DNS servers, while others operate as cache clusters. For increased uplink capacity and flexibility, orchestrators can also be installed as standalone units on electric poles, trees, or other outdoor fixtures within the same satellite spot beam.

[0133] These computing nodes are designed to operate across various satellite frequency bands, including L-band, S-band, Ka-band, and Ku-band. Orchestrators at the access layer use intelligent algorithms to determine both the number and selection of cache cluster nodes that should be assigned for serving SAT users within the same spot beam. All user internet requests are routed to these orchestrators, which then delegate the task to appropriate cache clusters nearby.

[0134] Each orchestrator maintains but not limited to a Distributed Hash Table (DHT) that contains detailed information about cache clusters and their active workload containers and the size of quality of the segments on the cache clusters. Based on this metadata, orchestrators select the optimal cache cluster nodes for segment downloads — whether for video content or files — ensuring minimal latency and efficient delivery. Cache clusters under the same spot beam act as worker nodes and directly serve SAT users with the requested content or services.

[0135] These cache clusters are not limited to video streaming; they are designed to host a wide range of workloads, including CRM platforms (e.g., Salesforce, Zendesk), eCommerce services, cloud productivity suites (e.g., Office 365, Google Workspace, Zoho Workplace), Al inference workloads, augmented and virtual reality (AR / VR) applications, gaming, and software updates. The access-layer orchestrators run on standard operating systems such as Linux or Windows and integrate orchestration platforms but not limited to likeAttorney Docket No.: 125512- 297358Kubemetes, enabling dynamic deployment of workloads across the distributed cache network. Meanwhile, the cache clusters themselves typically run lightweight OS environments and containerized workloads.

[0136] Each container within a cache cluster is dedicated to specific content or service provider — such as ABC News, CNN, YouTube, Netflix, Salesforce, etc. — ensuring logical separation and optimal performance. Orchestrators are also equipped with load-balancing capabilities using tools like but not limited to Nginx, Traefik, or HAProxy, and they serve as DNS servers via tools such as DNSmasq.

[0137] Workload deployment decisions are guided by real-time SAT user traffic patterns and can also involve coordination between neighboring orchestrators, whether within the same or across different spot beams, to maintain service continuity, redundancy, and optimal resource utilization.

[0138] 1. Low-cost, low-power computing units like Intel N150 / N100, RK3588 etc., will be hosted in Incentivized homes, outdoors like treetop, poles etc., with SAT modem and antenna serving as orchestrators and DNS servers. These computing units can be deployed as standalone with a higher uplink capacity under the same spot beam installed on electric poles, trees etc.,

[0139] 2. Low-cost, Low-power computing units like Intel N150 / N100, RK3588 etc., will be hosted in incentivized homes, outdoor deployments like treetop, poles etc., along with SAT modem and antenna serving as Cache Clusters.

[0140] 3. SAT modem hosting either access layer orchestrator or cache clusters will operate in either L-band, S-band, Ka-band or Ku-band.

[0141] 4. The access layer orchestrator will use appropriate algorithms to decide either how many cache cluster nodes or which nodes to be assigned to a SAT.

[0142] 5. SAT user’s Internet requests will be forwarded to the access layer orchestrators in the same spot beam, which will assign the cache clusters in the same spot beam.

[0143] 6. Access layer orchestrator will maintain a but not limited to DHT table with cache clusters and its workload container information based on which it will choose cache clusters nodes for downloading segments (video or files).

[0144] 7. Cache clusters hosted under the same spot beam will serve the SAT users.

[0145] 8. Cache clusters deployed in Incentivized homes serve as worker nodes to the access layer orchestrator.Attorney Docket No.: 125512- 297358

[0146] 9. The workloads on cache clusters can include any type of application, such as video streaming, CRM platforms (e.g., Salesforce, Zendesk), eCommerce services, cloud productivity suites (e.g., Office 365, Google Workspace, Zoho Workplace), Al inference workloads, augmented and virtual reality (AR / VR) applications, gaming, and software updates — all of which demand high availability and low latency.

[0147] 10. Access-layer orchestrators also run operating systems like Linux or Windows, and include orchestration platforms such as Kubemetes, which dynamically deploy workloads to the cache clusters.

[0148] 11. Cache clusters typically run lightweight operating systems such as Linux or Windows, and host lightweight containers.

[0149] 12. Each container running on the cache clusters serves a specific workload, such as ABC News, CNN, YouTube, Netflix, Salesforce, etc.

[0150] 13. These orchestrators include load balancers such as but not limited to Nginx, Traefik, or HAProxy.

[0151] 14. Cache clusters or orchestrators also serve as DNS servers, using tools like DNSmasq.

[0152] 15. Workloads are deployed on cache clusters based on SAT user traffic patterns, and may also be coordinated with neighboring orchestrators, which can belong to the same or different spot beams.

[0153] Video Streaming with MPRD for SAT Users:

[0154] For satellite (SAT) users, video content delivery is optimized through a distributed architecture involving cache clusters and intelligent workload orchestration. When a SAT user requests a video, the request is routed through a workload orchestrator located within the same satellite spot beam. The orchestrator identifies the appropriate workload container — such as one running a micro CDN server for YouTube — and assigns the task to one or more cache clusters installed at incentivized homes within the beam.

[0155] Frequently accessed videos are stored locally within these cache clusters, managed using caching algorithms like Least Recently Used (LRU), AI / ML-based or similar strategies to ensure optimal cache performance. Content from various providers — including YouTube, Netflix, and Disney + — is organized within dedicated workload containers or shared storage environments with proper isolation.

[0156] When a cache hit occurs in the cache clusters, the requested video is served directly from the local cache. On a cache miss, the content is retrieved from upstream CDN or edge nodes, multicast to both the requesting user and nearby cache clusters, and then locallyAttorney Docket No.: 125512- 297358stored for future use. This multicast approach ensures efficient content propagation and reduces repeated satellite uplink usage.

[0157] Users can receive video segments from a single cache cluster or from multiple clusters, depending on availability and system optimization. In scenarios where multiple cache clusters are involved, a receiver-driven Multipath Parallel Receiver-Driven (MPRD) algorithm enables the client to download video segments in parallel from different nodes. For example, if a user in a target home (as illustrated in FIG. 8) requests a video, the MPRD algorithm coordinates segment downloads from multiple incentivized homes, ensuring high throughput and seamless playback.

[0158] In an example implementation:

[0159] 1. SAT Users requesting video content will send their request to the workload orchestrator, which then forwards it to the container running the appropriate micro CDN server. For example, if a user requests a YouTube video, the request is forwarded to the workload orchestrator under the same spot beams, which intercepts the request and assigns it to the work nodes (cache clusters installed at incentivized homes) hosting the YouTube workload containers.

[0160] 2. Frequently viewed videos are stored in cache clusters. The micro CDN uses caching algorithms such as Least Recently Used (LRU), AI / ML-based algorithms or similar to manage and optimize video caching.

[0161] 3. Videos from each content provider — such as YouTube, Netflix, Disney+, etc. — are stored either in their respective workload containers or in shared storage with proper separation. These videos may reside in cache clusters or in centralized storage.

[0162] 4. On a cache hit in cache clusters, videos stored locally are delivered directly to the user. On a cache miss, the content is fetched from the CDN or edge nodes, then cached locally by enabling multicast feature on SAT onboarding and delivered to the user as well. Note: Multicast is transmitting video to multiple users - cache clusters and requesting user in this case.

[0163] 5. User-requested videos can be served from a single cache cluster or multiple cache clusters, depending on availability and system optimization.

[0164] 6. When a user is served from multiple cache clusters, the client uses a receiver-driven MPRD (Multipath Parallel Receiver-Driven) algorithm to download different segments from those clusters. For example, if a user from a target home in FIG. 8 requests a video, they may be served by four incentivized homes. The user at the target home uses the MPRD algorithm to receive all video segments, as described in the MPRD section.Attorney Docket No.: 125512- 297358

[0165] Cache Clusters for Mobile Internet Service Provider (Note: Mobile ISP deployment is similar to cache clusters for Satellite ISP except for the following)

[0166] Background. Cellular macro cells are high-powered base stations that provide wide-area wireless coverage for mobile users. Mounted on towers or rooftops, these base stations serve large areas — ranging from hundreds of meters in urban zones to several kilometers in rural regions. They form the core of mobile networks, handling user connectivity for voice, data, and messaging over licensed spectrum bands such as LTE and 5G. However, macro cells alone can struggle with congestion, latency, and spectrum limitations in high-density environments.

[0167] To enhance network performance and extend high-speed access to end-users, 5G Customer Premise Equipment (CPE) is increasingly used at homes and businesses. 5G CPE devices connect directly to the nearest 5G macro or small cell and provide Wi-Fi or wired internet access inside the premises. These devices enable fixed wireless access (FWA), offering fiber-like speeds without the need for physical cabling. Combined with edge computing and distributed caching, 5G CPE helps offload traffic from macro cells, reduce latency, and improve the quality of service for both residential and enterprise users.

[0168] In this solution, similar to SAT solution, the following may be added or amended:

[0169] 1. Multiple Industrial Grade CPE (Customer Premise Equipment) integrated with cache clusters are deployed for every Macro Cell as shown in FIG. 10.

[0170] 2. Access Layer orchestrator will also be integrated along with Industrial Garde CPE and deployed under the same macrocell.

[0171] 3. Similar to SAT users, mobile users will also be directed to access layer orchestrator, which will forward it to the appropriate container running appropriate micro CDN.

[0172] 4. An uplink classifier is used to direct packets within the macro cell, which will be used to redirect mobile users to access layer orchestrator and cache clusters integrated in Incentivized homes.

[0173] 5. The rest of the points 1-6 mentioned above can be brought in from the SAT solution.

[0174] FIG. 10 depicts a system architecture 1000 for cache clusters and access layer orchestrator for a mobile network such as 5G or Long-Term Evolution (LTE) network, in accordance with various embodiments. A base station 1002 communicates with a workflow orchestrator 1004, nodes 1010, 1020 and 1030, a portable user device 1040 such as a smartAttorney Docket No.: 125512- 297358phone or tablet, and a CDN / edge node 1050. The node 1010 includes CPE 1011 and cache devices 1012-1013, the node 1020 includes CPE 1021 and cache devices 1022-1023, and the node 1030 includes CPE 1031 and cache devices 1032-1033. The nodes 1010, 1020, 1030 and the CDN / edge node 1050 can upload data to the base station via wireless uplink paths 1014-1017, respectively, and the user device can download corresponding data in a wireless downlink path 1041. The base station can repackage the different data streams into a single stream for download by the user device.

[0175] In this example, the workflow orchestrator 1004 is a separate device from the nodes. Optionally, the workflow orchestrator can be incorporated into one of the nodes.

[0176] An industrial grade CPE is connected to the macro-cell. Incentivized nodes are distributed with the macro cell site locations.

[0177] Multipath Reverse Segment Download (MPRD) algorithm for segment downloads from multiple nodes or cache clusters.

[0178] See further discussion below.

[0179] I. Further discussion

[0180] As mentioned, over the past few years, the rapid proliferation of Internet-connected devices and the surge in data traffic demand have posed significant challenges to traditional centralized content delivery architectures, including reduced reliability and high costs. To address these issues, PCDNs based on decentralized architectures have emerged, enabling Internet-connected devices to share data directly with one another.

[0181] Decentralized data sharing has roots in early technologies such as USENET, which was used for news dissemination, and later evolved to support applications like music sharing (e.g., Napster) and file distribution (e.g., BitTorrent). In recent years, the growing demand for video-based applications — particularly short-lived videos — and the widespread adoption of mobile devices have prompted video content providers, such as ByteDance, to adopt decentralized PCDN architectures to enhance scalability and reduce cost of content delivery.

[0182] Traditional CDNs rely on large-scale deployment of dedicated servers, which costs hundreds of millions of dollars annually to deploy and maintain. These costs are also expected to significantly increase with the emergence of new content (e.g., mixed reality) due to the need to provision more capacity. PCDNs, on the other hand, leverage under-utilized resources at the network edge to store and distribute data, resulting in much lower cost to serve each user’s data. PCDNs can be either deployed as a fully decentralized replacement of CDNs or deployed alongside traditional CDNs to create a hybrid architecture.Attorney Docket No.: 125512- 297358

[0183] In a PCDN architecture, edge devices, including set-top boxes, smart home gadgets, and access points function as peer-to-peer nodes. These devices offer resource elasticity, allowing edge devices to join or leave the network dynamically and enabling content providers to scale resources up or down based on demand. While decentralized architectures address certain limitations of centralized systems, they also introduce new challenges. These include less compute and storage resources and less stable communication links compared to CDN architectures. To provide an equivalent user experience, PCDN architecture can use multiple edge devices to distribute data in parallel to each client. This parallel data transmission can increase the communication bandwidth, reduce latency, and increase robustness to link or server failures.

[0184] Over the last two decades, multipath transport protocols such as Multipath TCP (MPTCP) and Multipath QUIC (MPQUIC) have been developed to enable communication between a single client and a single server using multiple network paths simultaneously. These transport protocols can be adapted and deployed in PCDN architectures, enabling communication between multiple servers and a single client. However, the heterogeneity in path data rates and packet drop rates introduces challenges to existing multipath transport protocols (e.g., MPTCP / MPQUIC), such as Out-of-Order (000) packet delivery and Head-of-Line (HoL) blocking. Further, the change in architecture from a single server communicating with a single client (over multiple paths) to multiple servers communicating with a single client (over multiple paths) exacerbates these challenges. In fact, as demonstrated later, existing multipath transport protocols can get total throughput values even lower than those achieved with a single path.

[0185] To illustrate the HoL blocking and 000 packet delivery issues, consider the scenario illustrated in FIG. 3, where six data segments, numbered sequentially from 1 to 6, are transferred from three edge servers to a single client.

[0186] Here, different pieces of content come from different servers and paths. The heterogeneity in paths’ data rates and packet drop rate characteristics, can cause some parts of data (e.g., segments 3-6 in this figure) arrive earlier than other parts (e.g., segments 1 and 2). Since application requires packets to be delivered in order, this can cause HoL blocking, degrading overall throughput and client QoE.

[0187] Each edge server is assigned two segments: edge server 1 transmits segments 1 and 2, edge server 2 handles segments 3 and 4, and edge server 3 transmits segments 5 and 6. If edge server 1 experiences delays or packet losses, while edge server 2 and edge server 3 deliver their data successfully and on time, the client receives segments 3 to 6 before segmentsAttorney Docket No.: 125512- 2973581 and 2. Despite the timely arrival of segments 3 to 6, the application cannot process them due to the missing data from edge server 1. Many applications such as music players and video players require data to be received in sequential order for proper playback. This scenario — where later data segments arrive but cannot be processed due to missing earlier segments — results in OOO data arrival and leads to HoL blocking. Consequently, this phenomenon degrades throughput and reduces the user’s Quality of Experience (QoE).

[0188] Such HoL blocking underscores challenges in decentralized multipath architectures, particularly in ensuring seamless data delivery from heterogeneous edge devices. Overcoming these issues is essential to enhance the efficiency and reliability of decentralized systems, especially as demand grows for high-performance content delivery in applications like video streaming and immersive experiences.

[0189] In accordance with embodiments herein, HoL blocking and 000 packet delivery limitations stem from fundamental issues in existing multipath protocols like MPTCP, where components such as schedulers, congestion control, and retransmission queues at the multipath level conflict with similar mechanisms at the single-path level. To address these challenges, MPRD (Multipath Parallel Reverse Segment Download) is introduced, a novel distributed transport protocol specifically designed for PCDN architectures. MPRD builds on single-path TCP (SPTCP) and serves as a replacement for MPTCP in PCDN networks. Specifically, the solutions include the following:

[0190] MPRD Design: In accordance with certain embodiments, there is developed a new multipath transport protocol, MPRD, designed specifically for PCDNs. MPRD is a receiver-driven solution in which the client establishes multiple SPTCP connections and coordinates the starting point and transmission window offset at each server to parallelize segment downloads. The selection of edge servers is guided by information from a tracker entity and the path characteristics to each edge server. MPRD eliminates HoL blocking and 000 packet delivery observed in existing multipath schedulers, incurs minimal communication overhead, and achieves a total throughput close to the sum of the throughput of each individual path.

[0191] Evaluation: Extensive experiments have been conducted to evaluate MPRD across a variety of applications, path characteristics (data rate, loss), and numbers of clients and servers. Compared to the state-of-the-art, MPRD increases the aggregate throughput by a factor of 4-30x. MPRD reduces the load on the CDN up to 90%, when the CDN is augmented by a PCDN. For video applications, MPRD significantly reduces the percentage of pauses (stalls) and improves the PSNR, which is a key metric for visual quality, by up to 35 dB.Attorney Docket No.: 125512- 297358

[0192] The discussion below is organized as follows: Sections II and III cover the background and motivation. Section IV details the MPRD design. Implementation and evaluation are discussed in Sections V and VI, respectively. Related work and conclusions are presented in Sections VII and VIII.

[0193] II. Background

[0194] CDN. FIG. 11 depicts how a CDN can use a hierarchical structure 1100 with an origin or core server 1110 at the root, regional data centers 1120 in the middle, and edge servers 1130 close to the clients / end users 1140, in accordance with various embodiments.

[0195] A CDN (depicted in FIG. 11) operates on a hierarchical client-server architecture, with clusters of servers distributed across various geographical regions. These clusters deliver content, such as videos and files, to edge servers located closer to users, reducing latency and improving performance. CDNs use the Domain Name System (DNS) to minimize communication delays by directing users to the nearest edge server, ensuring load balancing and efficient resource utilization. Providers often scale server clusters and integrate dedicated ISP networks to maintain high-speed, low-latency communication. Prominent CDN providers, such as Akamai and Limelight, either deploy edge servers globally or rely on centralized data centers interfacing with ISPs. However, the high infrastructure costs required to scale CDNs and meet the growing bandwidth demands of Internet users remain a significant challenge.

[0196] PCDN. FIG. 12 depicts how a PCDN can employ a decentralized architecture 1200, operating standalone or in hybrid mode to reduce CDN load, in accordance with various embodiments. The architecture can include a set-top box, a smart TV, client devices, routers, and edge servers.

[0197] A PCDN is a decentralized system architecture where peers (e.g., edge devices) function as both content providers and consumers. In a PCDN, each peer can contribute to content storage, forwarding, and / or consumption. Unlike the DNS-based redirection employed in CDNs, PCDNs use a tracker entity to connect clients with peers that store the requested content. A PCDN can either be deployed in isolation as a replacement of a CDN or alongside a CDN to create a hybrid CDN-PCDN architecture. One key benefit of a hybrid architecture is to reduce the traffic load on the original CDN server.

[0198] Tracker. FIG. 13 depicts an architecture 1300 in which a tracker 1310 lists edge servers and their stored content, guiding clients (edge devices 1330) to the appropriate servers (PCDN users 1320) for downloads, in accordance with various embodiments. The tracker 1310Attorney Docket No.: 125512- 297358communicates with PCDN users 1320 and edge devices 1330, where the PCDN users can provide parallel downloads to the edge devices.

[0199] A tracker serves as a key component that facilitates efficient content delivery by managing peer discovery and connection setup. Unlike traditional CDNs that rely on DNSbased redirection, the tracker operates as a coordinating entity, maintaining a directory of peers and the content they store. When a client requests specific content, the tracker identifies and provides a list of suitable peers that can serve the request, optimizing selection based on factors such as proximity, bandwidth, and network load. By helping establish connections between clients and peers, the tracker ensures seamless data retrieval while supporting load balancing and efficient resource utilization. This functionality makes the tracker a cornerstone of decentralized architectures, enabling scalable and cost-effective content delivery.

[0200] User Interactions in PCDNs. The workflow of a PCDN revolves around the collaborative sharing of content among peers, such as edge devices. When a client requests specific content, it first communicates with a tracker entity, which maintains a directory of available peers and their stored content. The tracker identifies suitable peers based on criteria like proximity, bandwidth, and availability, then provides the client with a list of peers capable of serving the content. The client establishes direct connections with these peers and downloads the content from one or multiple sources. Throughout this process, peers may also cache the content they download, making it available to other clients and reducing the load on the original content source.

[0201] Multipath Transport Schedulers. Multipath transport for PCDNs enables clients to leverage multiple network paths simultaneously to improve data throughput, reduce latency, and enhance reliability. In this context, multipath transport protocols such as MPTCP can facilitate communication between clients and multiple peers or servers. To achieve optimal performance, these protocols rely on schedulers to determine how traffic is distributed across paths. Among the widely adopted MPTCP schedulers, minRTT focuses on optimizing performance by prioritizing data transmission over subflows with the lowest Round-Trip Time (RTT). It retransmits blocked segments via the fastest available path, penalizing subflows responsible for delays. While this approach ensures lower latency, it can result in underutilization of bursty network paths, limiting its ability to achieve optimal capacity aggregation. To address these issues, the BLEST scheduler was introduced to enhance MPTCP performance over heterogeneous paths. BLEST actively monitors the send window and minimizes delays caused by insufficient buffer space, ensuring faster subflows are not unnecessarily idle.

[0202] III, MotivationAttorney Docket No.: 125512- 297358

[0203] Preliminary experiments were conducted to highlight the poor performance of existing MPTCP schedulers in PCDN architectures. The experimental topology consists of two servers: a CDN server and an edge server, both hosted on a single Linux desktop machine. A single client, instantiated on a separate Linux machine, is connected to both servers. Details of the system implementation are provided in Section V.

[0204] FIG. 14 depicts a bar chart showing the throughput of each server to the client, both in isolation and when serving the client jointly. Without packet loss or delay, the CDN server achieves a data rate of approximately 10 Mbps, and the edge server achieves 4 Mbps (leftmost bar). When using MPTCP’ s minRTT scheduler for multipath download, the aggregate throughput is about 5% lower than the sum of the individual server throughputs. However, introducing packet loss or delay to the edge server using Linux TC reveals a significant impact on multipath download with minRTT scheduler.

[0205] Linux TC (Traffic Control) is a powerful command-line utility in the iproute2 package used to configure the Linux kernel’s packet scheduler. It manages network bandwidth by shaping (limiting rate), scheduling (prioritizing), and policing (dropping) traffic on network interfaces to ensure Quality of Service (QoS).

[0206] While packet loss or delay has visible effect on the edge server’s standalone data rate, the impact on aggregate throughput with minRTT is substantial: a 10% packet loss rate reduces throughput by 79%, and a 200 ms delay reduces throughput by 60%. Repeating the experiments with the BLEST scheduler shows only minor improvements in throughput. This is because existing MPTCP schedulers are highly susceptible to 000 packet delivery and HoL blocking, which degrade throughput.

[0207] The reduction in transport layer throughput directly impacts client QoE. For instance, FIG. 15 depicts a plot of Y-channel Peak Signal-to-Noise Ratio (Y-PSNR), a visual quality metric, versus encoding data rate, for an example video entitled “Big Buck Bunny,” in accordance with various embodiments. A drop in video bitrate from 4 Mbps to 1 Mbps reduces the PSNR from approximately 46 dB to 36 dB, indicating a significant decline in visual quality. This highlights the need for robust multipath transport solutions capable of maintaining high throughput despite packet loss or delay, conditions commonly encountered in both wireless networks (e.g., due to mobility or interference) and wired networks (e.g., due to router congestion).

[0208] IV. Design of MPRD

[0209] In this section, the design of MPRD is introduced, starting with an overview of its architecture, followed by a detailed explanation of the protocol.Attorney Docket No.: 125512- 297358

[0210] IV.A. Overall Architecture

[0211] Downloading content concurrently from multiple servers is an effective way to enhance throughput. However, ensuring sequential data delivery in the face of varying link data rates and packet loss presents a significant challenge. For example, consider two servers, Edge Server 1 and Edge Server 2, tasked with delivering six data segments labeled 1 through 6. A naive strategy might assign Edge Server 1 to send segments 1-3 and Edge Server 2 to send segments 4-6. In this case, any delay or packet loss from Edge Server 1 would make segments from Edge Server 2 unusable until segments 1-3 arrive. This problem, known as the HoL blocking issue, reduces effective throughput to the slower server’s rate, rather than achieving the combined throughput of both servers.

[0212] To address the HoL blocking limitations, a bidirectional downloading mechanism is proposed. In this approach, Edge Server 1 leverages SPTCP to transmit data sequentially in ascending order (1 to 6), while Edge Server 2 leverages another SPTCP to transmit in descending order (6 to 1). This dual-direction strategy allows the streams to converge, enabling sequential assembly of data at the client and effectively combining the throughput of both servers. For instance, if Edge Server 1 encounters delays transmitting segment 2, Edge Server 2 can finish segments 6 to 4 and then transmit segments 3 and 2, ensuring seamless delivery.

[0213] This approach maximizes throughput by leveraging the full capacity of both servers, despite individual uncertainties. Additionally, it enhances reliability and optimizes resource utilization, making it a robust solution for content delivery in PCDNs. Building on this principle, the mechanism is extended to multiple download sources and introduce the MPRD scheduler, which operates over single-path TCP.

[0214] The MPRD scheduler facilitates simultaneous download of data segments in both forward and reverse directions, using specific segment offsets to preserve order. The client first contacts the tracker to retrieve a list of servers hosting the requested content. MPRD then assigns the most reliable server (e.g., based on historical data) to transmit segments in the forward direction, while other servers handle reverse direction transmissions. To ensure orderly data delivery, the client assigns a distinct segment offset to each edge server.

[0215] Through this coordinated approach, MPRD eliminates the delays caused by 000 packets, improving throughput and minimizing the buffering overhead at the client. By aligning segment downloads in multiple directions, MPRD optimizes data flow and balances network load across multiple servers, making it a powerful solution to meet the demands of real-time applications such as streaming and live content delivery.Attorney Docket No.: 125512- 297358

[0216] MPRD emphasizes using the most reliable server to transmit data in the forward direction, while other servers handle reverse-direction transmissions. When segments from both directions converge, the client communicates updated segment offsets to all servers. The order of reverse-transmitted segments and the sequence of participating servers are determined based on their historical performance metrics.

[0217] For example, consider the scenario illustrated in FIG. 4 with one CDN server and two edge servers, all of which have access to data segments 1 through 6. The CDN server begins transmitting in the forward direction, starting with segment 1, while Edge Server 1 and Edge Server 2 transmit in the reverse direction, starting with segments 4 and 6, respectively. This reverse transmission strategy allows each edge server to operate at its own data rate.

[0218] When forward and reverse transmissions overlap, the client automatically detects the convergence point and adjusts segment offsets to maintain seamless delivery. The CDN server continues to deliver segments sequentially at its data rate, supplemented by the reverse-transmitted segments from the edge servers. These reverse segments effectively accelerate the overall throughput by augmenting the forward data flow.

[0219] This multipath transmission strategy eliminates HoL blocking, reduces retransmission overhead, and significantly increases total throughput.

[0220] IV.B, MPRD Detailed Design

[0221] For ease of discussion, consider a hybrid CDN+PCDN architecture composed of a CDN server and some edge servers.

[0222] In MPRD, the client would simultaneously request the data (e.g., files or videos) from the CDN server and quires the access layer orchestrator for a list of eligible edge servers installed at the incentivized homes. The tracker responds with a list of proximate edge servers capable of serving the requested content.

[0223] Unlike conventional vertical buffering mechanisms, MPRD on the client uses a horizontal buffer, known as the transmission window (TW), with a predefined capacity of N megabytes. All servers load the same data segments in their transmission window but end up transmitting different sets of segments. The client opens a similar window on its side (referred to as reception window) and populates it as segments arrive. Each data segment within each transmit window buffer is assigned the same sequence numbers, ensuring orderly data management.

[0224] At the outset, the MPRD scheduler sends a synchronized message to the CDN server and selected edge servers. This message includes essential details such as the requested content, the start sequence number, the offset, and the transmission direction for each server.Attorney Docket No.: 125512- 297358The start sequence number specifies the first segment in the requested data, while the offset indicates the position from which the server should begin transmitting segments. The direction parameter determines whether data transmission proceeds in ascending (forward) or descending (reverse) order.

[0225] The forward server, responsible for transmitting data in ascending order, is ideally designated as the CDN server. However, if the CDN server is unavailable, the edge server with the highest throughput is selected as the forward server. Edge servers are then sorted based on their estimated throughput and positioned within the transmission window. The initial throughput estimate can be determined by W / RTT, where W denotes the congestion window size and RTT is the roundtrip time between the server and client.

[0226] Each edge server is assigned a transmission window offset, computed using Eqs. 1 and 2. Here, K is the total number of edge servers, Rk denotes the throughput of edge server k, and RO denotes the CDN throughput. In Eq. 1, the estimated number of segments to be sent by each server is first determined. These estimates are then aggregated to compute the transmission window offset for each edge server (Eq. 2). See FIG. 16.

[0227] FIG. 16 depicts equations to compute the transmission window offset for each edge server, in accordance with various embodiments.

[0228] The transmission window offset determines the starting sequence number for each edge server’s transmission. While the forward server transmits data sequentially in ascending order and does not have an offset, all other edge devices transmit data in descending order, starting from their assigned transmission window offsets. Initially, MPRD distributes edge servers across the transmission window based on estimated throughput. However, this distribution may dynamically adjust across different transmission windows to better align with the throughput and reliability of each individual server over time.

[0229] MPRD assigns TWoffset (starting sequence numbers) and transmission directions (either forward or reverse) to all servers. Upon receiving these instructions, servers begin populating the client’s reception window buffer according to their assigned starting sequence numbers and directions. The MPRD scheduler continuously monitors data flow within the client’s reception window, employing pointers to track progress for each server. Specifically, the forward server is tracked using fwd ptr, while reverse servers are monitored using rev_ptn, rev_ptr2, ..., rev_ptrK. FIG. 17A shows these pointers on the client’s reception window.Attorney Docket No.: 125512- 297358

[0230] FIG. 17A depicts an initial placement of forward and reverse pointers at the client reception window, where the dashed parts represent the segments transmitted by the associated server, in accordance with various embodiments.

[0231] Initially, these pointers are positioned at specified offsets, as communicated to the servers. As data transmission progresses, the pointers increment or decrement their positions in the client’s reception window, based on received sequence numbers. Over time, some pointers would start to overlap and cross each other. The scheduler identifies crossover points — positions where two or more pointers intersect. Each pointer has defined crossover conditions: for instance, fwd_ptr encounters a crossover when it aligns with any rev_ptr. Similarly, for reverse pointers, rev_pti'k experiences a crossover when it reaches the initial position of rev_ptrk-i. The crossover points are based on the speed of segment transmissions from each server.

[0232] Managing crossover points is important for achieving high throughput in PCDN networks with heterogeneous server data rates. In MPRD, once a crossover is detected, the corresponding servers halt their transmissions and enter an idle state. To optimize resource utilization, idle pointers are reassigned by the client to manage undelivered data chunks with lower sequence numbers within the transmission window. The detailed mechanics of these crossovers are discussed next.

[0233] Forward and Reverse Pointer Crossover. When the forward server and reverse server 1 finish transmitting segments in TWoffsetl, MPRD migrates the fwd ptr and rev_ptrl to assist in retrieving unreceived segments originally allocated to rev_ptr2. This migration prioritizes rev_ptr2 over other reverse pointers due to the urgency of receiving lower-sequence-numbered data compared to higher- sequence-numbered data. (Some applications such as live streaming and video playback may be able to process orderly received lower-sequence-numbered segments without having to wait for the entire client reception window to be filled.) Initially, segments within TWoffset2 have been assigned to rev_ptr2 for reverse transmission. However, when portions of TWoffset2 remain unreceived, the MPRD scheduler redistributes these remaining segments to fwd_ptr and rev ptr 1 , as depicted in FIG. 17B.

[0234] FIG. 17B depicts how, when forward and reverse pointer 1 cross over (i.e., finish transmitting the segments between them), MPRD reassigns them to help with transmitting segments in T Woffset2, and the blocks 1710 and 1720 show the segments received by the client, in accordance with various embodiments.Attorney Docket No.: 125512- 297358

[0235] The new segment offset for rev_ptrl is determined based on Eq. 2. For smaller-sized unreceived segments, rev„ptrl is positioned near the midpoint of the remaining data to balance the transmission load efficiently.

[0236] Reverse Pointer Crossover. While fwd_ptr and rev ptr I are generally expected to complete their transmissions earlier than reverse pointers assigned to other servers, link rate and loss variations across servers can lead to different outcomes. In some cases, one or more servers associated with reverse pointers may complete their tasks ahead of schedule, leaving them idle. To address this, the MPRD scheduler reallocates these idle servers to accelerate the retrieval of unreceived segments associated with fwcLptr and rev„ptrl. This reassignment prioritizes lower- sequence-numbered data to ensure efficient data transmission. Idle reverse pointers are positioned between fwd.ptr and rev. ptr 1, as determined by Eq. 2. For smaller-sized unreceived TW offsets, the new reverse pointer is placed closer to the midpoint between fwcLptr and rev_ptrl, as illustrated in FIG. 17C.

[0237] FIG. 17C depicts how, as reverse pointer 2 finishes transmitting its segments, MPRD reassigns it to be in between the forward pointer and reverse pointer 1, in accordance with various embodiments. This approach optimizes the distribution of resources, ensuring that idle servers contribute effectively to minimizing delays in data retrieval.

[0238] MPRD Total Throughput is Close to the Summation of Individual Throughput Values. To maximize download efficiency from multiple servers, it is essential to use the full capacity of individual servers, while avoiding inter-dependencies that can cause HoE blocking. MPRD achieves this by serving the transmission window from both forward and reverse directions, and continuously monitoring and reassigning servers as servers finish their tasks. This can, however, sometimes cause redundant packet transmissions at the transport layer. For example, suppose there are two segments in the transmission window and two servers. Suppose forward server has sent segment 1 and reverse server has sent segment 2. Now suppose the client receives segment 1 but not segment 2. In this case, forward server will send segment 2 irrespective of the transmission from reverse server. Delayed arrival of segment 2 from reverse server will result in duplication. Duplicate packets will be detected and dropped at the client transport layer, however, this results in a total MPRD throughput that is slightly less than individual data rates as shown through experiments described herein.

[0239] Transmission / Reception Window Update. As the client reception window begins to fill, MPRD ensures that data is delivered to the application either as a complete set or incrementally, based on application requirements. Once the reception window reaches capacity, the entire client reception window is cleared (handed to the application layer).Attorney Docket No.: 125512- 297358Similarly, each server refreshes its transmission window with the set of new segments that should be delivered to the client. Further, MPRD evaluates the throughput of each server and dynamically reassigns them as forward or reverse servers, as described in the design section. This periodic reassessment optimizes resource allocation for subsequent transmission windows, ensuring sustained performance and efficient data delivery.

[0240] Congestion Control. MPRD supports both coupled and uncoupled congestion control. The current implementation of MPRD is based on uncoupled congestion control, where each SPTCP operates independent of other SPTCPs. Uncoupled congestion control can sometimes lead to unfairness and over the past several years, several works have been proposed to extend MPTCP to support coupled congestion control for improved fairness. These approaches can be integrated into MPRD by regulating the amount of traffic served by each server and / or designing new pointer distribution mechanisms within the transmission window.

[0241] V, Implementation

[0242] In this section, an implementation is detailed, which comprises two servers connected via multiple Ethernet interfaces, along with software to emulate (virtualize) various configurations of servers, clients, and multipath characteristics (e.g., delay, packet loss rate, data rate). Specifically, generic client and server software are tailored for the PCDN architecture. Three multipath schedulers and transport protocols are then implemented: MPTCP minRTT, P-Scheduler, and MPRD. The implementation is based on Linux Ubuntu 24.04 LTS. Additionally, custom scripts and libraries were created to process and generate diverse types of data (e.g., file downloads, video processing files).

[0243] Hardware Setup. The experimental setup, depicted in FIG. 18, comprising two Dell desktops equipped with Intel Core i7 processors (12 cores, 3.7 GHz), 32 GB of RAM, and running Ubuntu 22.04 LTS.

[0244] FIG. 18 depicts client and server processes on two separate machines 1810 and 1820, in accordance with various embodiments. The machines were interconnected through up to five dedicated Ethernet connections using USB-to-Ethernet adapters, each supporting speeds of up to 1 Gbps. These dedicated Ethernet interfaces provided precise control over bandwidth (data rate) and packet loss on each connection, enabling the emulation of real- world PCDN scenarios under varying resource constraints. While the entire setup could have been emulated on a single Linux machine using a loopback interface, this approach posed limitations: bandwidth would be shared among servers, and it would be difficult to control bandwidth or induce packet loss on individual connections. To address these challenges, the setup was extended to two Linux machines with dedicated Ethernet interfaces.Attorney Docket No.: 125512- 297358

[0245] Client and Server Software. Generic client and server software was developed for PCDNs using C++ in Linux environments. The PCDN client software supports configurations for minRTT, P-Scheduler, and MPRD, offering the flexibility to scale to any number of clients and servers. In this architecture, each client can spawn a configurable number of processes, with each process independently receiving data from one of the PCDN servers (e.g., one CDN and one or more edge servers), as illustrated in FIG. 19.

[0246] FIG. 19 depicts how Linux TC can control bandwidth, delay, and loss for each path, in accordance with various embodiments.

[0247] This design choice — dedicating a process to each server rather than using a single process to handle multiple servers — was made to ensure better resource allocation and higher throughput between clients and servers. On the server side, the architecture similarly supports a configurable number of processes, each functioning as an independent server (CDN or an edge server). Unique server port numbers and IP addresses are used to pair each clientserver connection, enabling efficient, isolated communication with configurable bandwidth and packet loss for each pair.

[0248] Bandwidth and Loss Control. To induce packet loss and control bandwidth, the TC and Wondershaper tools were used in Linux. TC was used to introduce packet loss on specific interfaces between the client and server, while Wondershaper regulated bandwidth, allowing us to emulate PCDN edge or CDN servers under varying resource and bandwidth constraints. However, these tools did not perform as expected on virtual interfaces or loopback connections, making it difficult to achieve dedicated loss and bandwidth control. This limitation necessitated the use of two physical servers with dedicated Ethernet interfaces to ensure precise and reliable parameter configuration.

[0249] PCDN System Framework and Multipath Transport Implementation. In the implementation, each client process establishes a bidirectional socket connection with the server. One socket is dedicated to downloading the requested content from the server, as specified by the client-side scheduler, while the other socket handles metadata communication. This metadata includes information such as the content (e.g., video) ID, start sequence number offset, and download direction in the case of MPRD. In the performance evaluation, up to five paths (client- server process pairs) between a server and a client were considered, as depicted in FIG. 19. Three multipath transport protocols and schedulers were then implemented:

[0250] L minRTT: The minRTT scheduler was integrated into the PCDN system software based on the minRTT approach from MPTCP. In this approach, the client-side scheduler prioritizes requests to servers with the lowest RTT, ensuring that the congestionAttorney Docket No.: 125512- 297358window of the PCDN connection is filled by these servers first before addressing those with higher RTTs. In the controlled environment, RTT values remain constant unless packet losses are deliberately introduced. The server is pre-configured with the RTT values for each clientserver path and uses this information to prioritize filling the congestion window of the lowest RTT connections first, followed by higher RTT connections, in alignment with the minRTT strategy.

[0251] 2, P-Scheduler: The P-Scheduler mechanism was incorporated into the PCDN system design. Each server is assigned an offset calculated using parameters such as average transmission time, data rate, packet size, and retransmission timeout. At the start of the process, P-Scheduler computes the offset for each client-server process pair based on its algorithm, and each server initiates its transmission according to the assigned offset. Once a server completes its transmission within the specified offset, P-Scheduler dynamically reassigns a new offset to the client-server process pair, ensuring efficient and orderly data transfers across connections.

[0252] 3, MPRD: The multipath transport protocol architecture and scheduling follow the design outlined in Section IV. Specifically, the scheduler assigns sequence numbers and forward / reverse directions to all client-server process pairs. Each client process reports successfully received data to the MPRD scheduler using sequence numbers. The scheduler continuously monitors the transmission window for crossovers, as described in Section IV. Upon detecting a crossover, the MPRD scheduler promptly re-assigns sequence numbers to the completed client-server process pairs, ensuring efficient data flow and seamless coordination across the different paths.

[0253] VI. Results and Discussion

[0254] In this section, the results of extensive performance evaluation are reported. MPRD’s performance is analyzed in terms of throughput aggregation in both high and low data rate regions. Next, the reduction in CDN load is assessed when a PCDN is used to augment a CDN server. MPRD’s ability to enhance video performance is then evaluated using metrics such as PSNR and percentage of pauses (re-buffering). Finally, the time required to download a file is assessed, emulating a multipath BitTorrent application. In all evaluations, MPRD’s performance is compared against the default Linux scheduler (minRTT) and the state-of-the-art (P-Scheduler). Results from BLEST are omitted, as its performance consistently fell between minRTT and P-Scheduler.

[0255] Scenario. Ahybrid CDN+PCDN architecture is considered, where one or more edge PCDN servers augment the CDN server. The CDN server is characterized by a higherAttorney Docket No.: 125512- 297358data rate to the client compared to the edge servers. This scenario also captures a pure PCDN architecture, where one server achieves a higher data rate to the client than the other servers.

[0256] VI. A, Throughput Aggregation

[0257] Throughput Measurement Procedure. For all three multipath schemes examined in this study, each client process logs the received data, including sequence numbers and timestamps, into individual CSV files. These files are then post-processed. Sequence numbers are inserted into a priority queue for throughput estimation at 1- second intervals, based on the logged timestamps. The priority queue maintains an ascending order of sequence numbers, and the total in-sequence data in the queue represents the number of packets received within a given second. In this study, a data packet size of 1400 bytes was used in all experiments. Throughput was calculated by multiplying the number of in-sequence packets in the priority queue per second by the packet size.

[0258] Throughput Statistics. Video service providers, such as ByteDance, report that 80% of users in China stream videos at rates between 2-4 Mbps, with 1% utilizing up to 8 Mbps. Additionally, the average CDN speed in China is approximately 24 Mbps. In the USA, the average ISP downlink and uplink throughputs are 200 Mbps and 24 Mbps, respectively. For mobile networks, 5G and 4G in the USA achieve theoretical maximum downlink speeds of up to 2 Gbps, with an average downlink speed of 50 Mbps. Using these statistics, two scenarios were designed to evaluate throughput: one with low individual server data rates and another with high rates.

[0259] Specifically, tests were conducted using configurations with 2, 3, 4, and 5 total number of servers. In the low-throughput setting, server-to-client links were bandwidth-limited to 10 Mbps, 4 Mbps, 4 Mbps, 3 Mbps, and 1 Mbps in the five-server setup, with the CDN server providing the highest data rate (10 Mbps). This heterogeneous setup was designed to produce out-of-order packets, mimicking real-world network variability and enabling an assessment of each scheduler’ s effectiveness. For the high-throughput setting, bandwidths were set to 500 Mbps, 350 Mbps, 350 Mbps, 200 Mbps, and 200 Mbps for the respective server-to-client links. This configuration aimed to evaluate performance under conditions with higher bandwidth and less stringent link limitations.

[0260] Results are provided in FIGs. 20A-20F.

[0261] FIG. 20A depicts a bar chart of throughput across different schemes as a function of number of edge servers in a low throughput configuration, where no server experiences packet loss, in accordance with various embodiments.Attorney Docket No.: 125512- 297358

[0262] FIG. 20B depicts a bar chart of throughput results when edge server one experience 10% packet loss rate, in accordance with various embodiments.

[0263] FIG. 20C depicts a bar chart of CDN load as a function of number of servers (CDN server + zero to four edge servers), where no server experiences packet loss, in accordance with various embodiments.

[0264] FIG. 20D depicts a bar chart of throughput across different schemes in a high throughput configuration, where no server experiences packet loss, in accordance with various embodiments.

[0265] FIG. 20E depicts a bar chart of throughput results when edge server one experiences 10% packet loss rate, in accordance with various embodiments.

[0266] FIG. 20F depicts a bar chart of CDN load as a function of number of servers (CDN server + zero to four edge servers), where edge server one experience 10% packet loss rate, in accordance with various embodiments.

[0267] All three schedulers — minRTT, P- Scheduler, and MPRD — were evaluated under two conditions: (1) without packet loss and (2) with a 10% packet loss introduced on edge server 1. For systems without packet loss, minRTT achieved about 80% of the total system throughput, P-Scheduler achieved about 85%, and MPRD demonstrated the best performance by achieving about 87-90% of the total system throughput across all configurations as shown in FIG. 20A (low throughput configuration) and FIG. 20D (high throughput configuration). Here, the first bar for a given x-axis value shows the throughput of each individual server in isolation. The second, third, and fourth bars show the throughput values of minRTT, P-Scheduler, and MPRD, respectively.

[0268] When a 10% packet loss was introduced to edge server one, MPRD consistently retained 75 to 85% of the total throughput across all configurations, while minRTT and P-Scheduler collapsed significantly during the 1, 2, 3, and 4 edge server configurations, achieving only 2-3% of the sum throughput as shown in FIG. 20B (low throughput configuration) and FIG. 20E (high throughput configuration). This collapse was attributed to the inefficiency of minRTT and P- scheduler to handle OOO packets and HoL blocking issues.

[0269] The results highlight that loss and jitter, common in real-world scenarios, can severely degrade the performance of min-RTT and P- Scheduler. On the other hand, MPRD demonstrated robustness and superior loss tolerance, maintaining higher throughput even under adverse conditions. These findings underscore MPRD’s effectiveness as a multipath transport solution for practical PCDN deployments, ensuring optimal QoE for users with varying number of servers.Attorney Docket No.: 125512- 297358

[0270] VLB. CDN Load Reduction

[0271] One of the primary challenges for content service providers using CDNs is addressing the demands of an ever-increasing user base and the escalating bandwidth requirements. Factors such as higher video frame rates and the growing adoption of emerging high-volume content formats (e.g., 4K or 8K video) place considerable strain on CDN infrastructure, both in terms of cost and scalability. A PCDN system can mitigate this bottleneck by offloading some of the content delivery responsibilities to edge servers. This distributed approach not only eases the load on central CDN servers but also helps reduce operator costs, making it an important solution for handling modem content delivery demands.

[0272] The effectiveness of a PCDN system in reducing CDN load depends on how well the edge servers offload traffic from the CDN, while maintaining the required user QoE. The load reduction capabilities of the three multipath systems noted herein were then evaluated.

[0273] CDN Load Measurement Setup. To quantify the reduction in load when a CDN server is augmented with edge PCDN servers, a series of measurements were conducted. The client downloaded data with sequential numbers from both the CDN and edge servers. During a 5-minute interval, the amount of data (in bytes) transferred from the CDN and each edge server to the client was logged. These measurements were taken at the interface between the transport and application layers because packets received 000 at the transport layer cannot be delivered to the application layer, and hence, do not contribute to application-layer throughput. The logged data was post-processed to calculate the CDN load. This load was determined as the ratio of the bytes received from the CDN server to the total bytes received from all servers.

[0274] CDN Load Reduction Characterization. FIG. 20C illustrates the CDN load as a function of the total number of servers. When there is only one server, it is solely the CDN server. In these experiments, no server was subjected to packet loss. Further, the server data rates remain the same as the high throughput setup in FIG. 20D. All three mechanisms were observed to achieve close CDN loads, regardless of the number of servers. This is because, in the absence of packet loss, all three schemes achieve close throughput values, as discussed in FIG. 20D. It was also observed that as the number of servers increases, the CDN load progressively reduces, demonstrating the efficacy of PCDNs to reduce CDN load.

[0275] FIG. 20F illustrates the CDN load as a function of the total number of servers, with edge server one experiencing a 10% packet loss rate. The results show that MPRD significantly outperforms both minRTT and P-Scheduler, achieving a significant reduction in CDN load. Additionally, as more servers are added to the PCDN system, MPRD consistentlyAttorney Docket No.: 125512- 297358maintains a substantially lower load on the CDN, demonstrating its effectiveness in mitigating the impact of packet loss.

[0276] VI.C. Video Performance

[0277] Video Data Set and Preparation. For the video streaming evaluations, a 5-minute clip was used from the widely used 3D animated short film Big Buck Bunny. The original video was encoded in H.264 format at a resolution of 3840x2160 and a frame rate of 60 fps. To prepare it for streaming, x264 via ffmpeg was employed to encode the video into four representations with bitrates of 1 Mbps, 2 Mbps, 4 Mbps, and 8 Mbps, with resolutions of 360p, 480p, and 1080p (for the two highest bitrates). The frame rate was fixed at 60 fps across all representations, with a uniform GOP (Group of Pictures) size of 60 frames, achieved by inserting keyframes at one-second intervals. During streaming, the client logged bytes received at one-second intervals and recorded any pauses when the received data fell short of the required amount for a single GOP. The percentage of pauses was calculated over the entire video duration. Additionally, video quality was assessed by computing PSNR values relative to the original video.

[0278] Video Performance in terms of PSNR and Stalled (Re-buffering) Times. A scenario is considered with one CDN with 10 Mbps data rate and one edge server with 4 Mbps data rate and 10% packet loss rate. FIGs. 21A and 21B show percentage of client pauses and PSNR versus video bitrate, respectively.

[0279] FIG. 21 A depicts a bar chart of the percentage of pauses for different data rates of Big Buck Bunny video, in accordance with various embodiments.

[0280] FIG. 2 IB depicts a bar chart of PSNR for different data rates of Big Buck Bunny video, in accordance with various embodiments.

[0281] FIG. 21C depicts a bar chart of file download time for different data rates of Big Buck Bunny video, in accordance with various embodiments.

[0282] From FIG. 21 A, minRTT and P-Scheduler experience significant re-buffering, with stall rates from 60% to 95%. Further, P-Scheduler only slightly improves video performance against min RTT in terms of pauses. MPRD consistently avoided re-buffering even for higher bitrate videos, highlighting its ability to deliver uninterrupted playback.

[0283] To further quantify the client QoE, the PSNR was measured for all three schedulers. A PSNR value of 55 dB is typically associated with good QoE at a given bitrate. For a 1 Mbps bitrate video, clients using minRTT and P-Scheduler experienced a significant degradation in QoE, with PSNR values dropping by 25 dB as shown in FIG. 21B. In contrast, MPRD achieved the target PSNR of 55 dB, delivering a high-quality user experience. At higherAttorney Docket No.: 125512- 297358bitrates of 2 Mbps, 4 Mbps, and 8 Mbps, minRTT and P-Scheduler further deteriorated, resulting in PSNR values of 25 dB, 20 dB, and 18 dB, respectively as shown in FIG. 2 IB. These results demonstrate the efficacy of MPRD in handling video applications across varying bitrates and network throughput conditions. MPRD not only ensures seamless playback but also maintains high PSNR, making it well-suited for real- world video delivery scenarios.

[0284] VI.D, File Download Time

[0285] Measurement Setup. For file download evaluations, files of predetermined sizes (200 and 400 MB) were transferred using each of the three multipath schedulers and protocols. The time taken to download files were recorded and analyzed to assess the performance of each method. Here, a topology composed of one CDN and one or two edge servers was considered. Further, edge server one has a 10% packet loss probability.

[0286] Download Time. FIG. 21C illustrates the download times as a function of file size and the number of edge servers. The results show that MPRD reduces download time by a factor of 20-30x, consistently outperforming both minRTT and P-Scheduler, regardless of file size or the number of edge servers. In other words, minRTT and P-Scheduler would require 20-30 times more servers to match MPRD’s download times.

[0287] This performance improvement is primarily due to the HoL blocking issues faced by minRTT and P-Scheduler when one server experiences a significantly higher packet loss rate compared to others. MPRD circumvents these limitations by employing bidirectional communication, which mitigates HoL blocking, increases total system throughput, and significantly reduces file download times.

[0288] FIG. 22 depicts an example communication architecture 2200 including a micro point of presence (M-PoP) 2210 in communication with an optical line termination (OLT) device 2220, in accordance with various embodiments. The M-PoP device includes a set of modules, e.g., module 2211 with connector 2212, that can communicate with the OLT device via a respective splitter in a set of splitters 2230. Each splitter can server, 32, 16 or 8 homes, for example. The OLT device also includes Ethernet ports 2221, 2222 and 2223 that can operate at 1G or 10G, for example. The XGB notation in this example refers to the Ethernet communication ports found on the LS Electric XGB series of Micro Programmable Logic Controllers (PLCs). These are industrial-grade ports designed for automation, machine control, and communicating with other devices on a factory floor.

[0289] A GPON port is a physical optical fiber interface located on an Optical Line Terminal (OLT) that provides high-speed, point-to-multipoint broadband connectivity (up to 2.4 Gbps) over passive fiber infrastructure.Attorney Docket No.: 125512- 297358

[0290] The M-PoPs enable deep edge computing in the access layer. A splitter loopback is connected to the high- availability (HA) unit (M-PoP 2210), which is tightly integrated with the OLT and splitters. The HA system can be located in a central office of a network, for example, such as depicted in FIG. 23. The architecture provides a platform for next-generation low latency applications.

[0291] FIG. 23 depicts an example communication architecture 2300 including the M-PoP 2210 at a central office 2310, in accordance with various embodiments. A number of OLTs 2311 and 2312 can be provided as well in the central office. The central office is coupled to a passive splitter 2320, which in turn is coupled to optical networking units (ONUs) 2330, 2331 and 2332. An ONU is a device in fiber-optic networks that converts optical signals from an ISP's Optical Line Terminal (OLT) into electrical signals for user devices. The outputs of the ONUs 2330 and 2331 are coupled to respective user premises, while the output of the ONU 2332 is coupled to the module 2211.

[0292] The architecture can be represented by a Passive Optical Network (PON) / Optical Distribution Network (ODN) 2340 coupled at one side to user premises and at another side to a M-PoP module 2350. A PON is a fiber-optic access technology that uses unpowered optical splitters to connect a single central office source (OLT) to multiple endusers (ONTs / ONUs) via a single fiber strand. An ODN is the passive fiber-optic infrastructure in a PON that connects the service provider's Optical Line Terminal (OLT) to the user's Optical Network Unit / Terminal (ONU / ONT). I

[0293] The M-PoP module in turn is coupled to a network 2360 which includes Kubemetes clusters with content provider services, as represented by cubes in the network. Kubemetes is an open-source container orchestration platform designed to automate the deployment, scaling, and management of containerized applications. A Kubemetes cluster is a set of node machines (physical or virtual) grouped together to run containerized applications, managed by a central control plane. It orchestrates containers across multiple environments, ensuring applications run with high availability, automatic scaling, and efficient resource utilization.

[0294] The architecture includes features such as a nearly zero-hop peer edge node within the access layer. The peer edge node serves content directly to subscribers within each PON. There is ultra-low latency such as less than 2-3 msec. Finally, throughput at the min (uplink, downlink) is about 10 GB for XGPON or 1 GB for GPON access bandwidth.

[0295] FIG. 24 depicts an example system architecture of the M-PoP 2210, in accordance with various embodiments. The M-PoP includes the modules, including theAttorney Docket No.: 125512- 297358example module 2211, a dual battery and redundant power supply unit (PSU) 2410, dual network switches 2415, and a CPU / storage / GPU 2420. The M-PoP is a high availability system. The access layer edge platform is deployed with the ISP network. Dual Uninterruptible Power Supply (UPS) and redundant power supplies can be provided for continuous operation. Dual 10G network switches can be provided for resilient connectivity. The architecture can supports 10-12 or more pluggable modules / nodes per HA system. Modular units are provided for CPU (Xeon / AMD), Storage, and GPU — with plug-and-play expansion to increase capacity. There is a scalable design with no single point of failure. Multitenant Kubemetes clusters can serve within each PON.

[0296] FIG. 25 depicts an example system architecture 2500 with M-PoP modules / nodes using Kubernetes clusters, in accordance with various embodiments. A module / node 2530 includes a number of pods, such as example pods 2531, a Kubelet 2532, and an operating system 2533. The Kubelet is the primary node-level agent in Kubemetes that runs on every worker node. It ensures containers are running in a Pod by enforcing the desired state defined by the control plane. It manages container lifecycles, reports status, and handles volume / network config. Similarly, a module / node 2540 includes a number of pods, such as example pods 2541, a Kubelet 2542, and an operating system 2543.

[0297] In Kubernetes clusters, each node hosts multiple containers (pods), pods are grouped into distributed Kubemetes clusters, and clusters host Over-The-Top (OTT), Gaming, and Al workloads, for example. OTT workloads refer to the streaming of video, audio, or messaging services directly over the internet, bypassing traditional cable, satellite, or broadcaster-managed networks. Additionally, every node in the PON-Peer Edge node has access to dedicated clusters, multi-tenant Kubemetes clusters can be provided, and contents such as gaming, OTT, augmented reality (AR) / VAR are hosted on the clusters

[0298] FIG. 26 depicts an example micro CDN architecture 2600 using Kubernetes pods and clusters, in accordance with various embodiments. Clusters 2610, 2620 and 2630 are depicted. The cubes in each cluster represent Kubemetes pods / containers. The Micro CDN core components can include a PON-Peer Edge node deployed at the access layer, multi-tenant Kubemetes clusters serving subscribers within each PON, and an Al - Cache Aware Adaptive Scheduler (Al-CaaS) for Intelligent content placement and traffic optimization. Other features include cluster and Mesh Network Discovery enabling co-ordination across M-PoPs, and Multipath Reverse Download for multipath parallel segment delivery with sum throughput across nodes.Attorney Docket No.: 125512- 297358

[0299] FIG. 27 depicts an example communication architecture 2700 in which the M-PoP 2210 is connected directly to an Ethernet port, in accordance with various embodiments. There are variations of deploying the M-PoP nodes at the access layer. Depending on the OLT and the number of supporting PONs, the architecture can vary. It can connect to the Ethernet ports of the OLT, for example. In particular, a module 2213 with connector 2214 is coupled to the Ethernet port 2222.

[0300] FIG. 28 depicts an example communication architecture 2800 in which the M-PoP 2210 is connected to an Ethernet port via a Virtual Local Area Network (VLAN) 2810, in accordance with various embodiments. In particular, the module 2213 with connector 2214 is coupled to the Ethernet port 2223 via the VLAN 2810.

[0301] Conclusion

[0302] Herein, MPRD, a multipath transport protocol, is designed to enhance content delivery in PCDN networks by leveraging multiple servers to concurrently transmit data to each client. MPRD, built on top of single-path TCP, addresses the problem of HoL blocking in PCDN systems by dynamically assigning transmission window pointers with optimized directions and placements for each server. MPRD is implemented alongside two other schedulers in the Linux userspace and demonstrated its significant performance benefits. The results show that when even one edge server experience packet loss, MPRD can achieve a 4-30x increase in total throughput, reduce CDN load by up to 85-90%, and improve PSNR visual quality by up to 35 dB.

[0303] Various operations may be described as multiple discrete actions or operations in turn, in a manner that is most helpful in understanding the claimed subject matter. However, the order of description should not be construed as to imply that these operations are necessarily order dependent. In particular, these operations may not be performed in the order of presentation. Operations described may be performed in a different order than the described embodiment. Various additional operations may be performed and / or described operations may be omitted in additional embodiments.

[0304] The terms “substantially,” “close,” “approximately,” “near,” and “about,” generally refer to being within + / - 10% of a target value. Unless otherwise specified the use of the ordinal adjectives “first,” “second,” and “third,” etc., to describe a common object, merely indicate that different instances of like objects are being referred to, and are not intended to imply that the objects so described must be in a given sequence, either temporally, spatially, in ranking or in any other manner.

[0305] For the purposes of the present disclosure, the phrases “A and / or B” and “A orAttorney Docket No.: 125512- 297358B” mean (A), (B), or (A and B). For the purposes of the present disclosure, the phrase “A, B, and / or C” means (A), (B), (C), (A and B), (A and C), (B and C), or (A, B, and C).

[0306] The description may use the phrases “in an embodiment,” or “in embodiments,” which may each refer to one or more of the same or different embodiments. Furthermore, the terms “comprising,” “including,” “having,” and the like, as used with respect to embodiments of the present disclosure, are synonymous.

[0307] As used herein, the term “circuitry” may refer to, be part of, or include an Application Specific Integrated Circuit (ASIC), an electronic circuit, a processor (shared, dedicated, or group), a combinational logic circuit, and / or other suitable hardware components that provide the described functionality. As used herein, “computer-implemented method” may refer to any method executed by one or more processors, a computer system having one or more processors, a mobile device such as a smartphone (which may include one or more processors), a tablet, a laptop computer, a set-top box, a gaming console, and so forth.

[0308] The terms “coupled,” “communicatively coupled,” along with derivatives thereof are used herein. The term “coupled” may mean two or more elements are in direct physical or electrical contact with one another, may mean that two or more elements indirectly contact each other but still cooperate or interact with each other, and / or may mean that one or more other elements are coupled or connected between the elements that are said to be coupled with each other. The term “directly coupled” may mean that two or more elements are in direct contact with one another. The term “communicatively coupled” may mean that two or more elements may be in contact with one another by a means of communication including through a wire or other interconnect connection, through a wireless communication channel or link, and / or the like.

[0309] Reference in the specification to "an embodiment," "one embodiment," "some embodiments," or "other embodiments" means that a particular feature, structure, or characteristic described in connection with the embodiments is included in at least some embodiments, but not necessarily all embodiments. The various appearances of "an embodiment," "one embodiment," or "some embodiments" are not necessarily all referring to the same embodiments. If the specification states a component, feature, structure, or characteristic "may," "might," or "could" be included, that particular component, feature, structure, or characteristic is not required to be included. If the specification or claim refers to "a" or "an" element, that does not mean there is only one of the elements. If the specification or claims refer to "an additional" element, that does not preclude there being more than one of the additional elements.Attorney Docket No.: 125512- 297358

[0310] Furthermore, the particular features, structures, functions, or characteristics may be combined in any suitable manner in one or more embodiments. For example, a first embodiment may be combined with a second embodiment anywhere the particular features, structures, functions, or characteristics associated with the two embodiments are not mutually exclusive.

[0311] While the disclosure has been described in conjunction with specific embodiments thereof, many alternatives, modifications and variations of such embodiments will be apparent to those of ordinary skill in the art in light of the foregoing description. The embodiments of the disclosure are intended to embrace all such alternatives, modifications, and variations as to fall within the broad scope of the appended claims.

[0312] An abstract is provided that will allow the reader to ascertain the nature and gist of the technical disclosure. The abstract is submitted with the understanding that it will not be used to limit the scope or meaning of the claims. The following claims are hereby incorporated into the detailed description, with each claim standing on its own as a separate embodiment.

Claims

Attorney Docket No.: 125512- 297358CLAIMSWhat is claimed is:

1. A decentralized Internet Service Provider (ISP) access layer-based Content Delivery Network (CDN), comprising:a plurality of low-power, low-cost cache nodes deployed at or near end-user premises within an access network, wherein the cache nodes are located in at least one of residential units, optical splitters, or at a headend of the access network;at least one workload orchestrator configured to identify an optimal subset of the cache nodes based on locality, congestion and performance history; anda client device configured to request content and concurrently retrieve content segments from the subset of cache nodes using a receiver-driven coordination protocol, wherein the coordination protocol schedules segment downloads in forward and reverse directions to aggregate throughput and minimize head-of-line blocking.

2. The CDN of claim 1, wherein the cache nodes operate at a low power consumption of less than 15-30 watts.

3. The CDN of claim 1, wherein the coordination protocol is a Multi-Path Reverse and Forward Direction (MPRD) protocol.

4. The CDN of claim 3, wherein the client determines a download direction for each cache node and adjusts segment offsets dynamically based on node performance.

5. The CDN of claim 1, wherein the cache nodes are hosted in incentivized residential homes.

6. The CDN of claim 1, wherein the workload orchestrator is configured to deploy micro CDN or edge computing workloads including at least one of video streaming, Artificial Intelligence (Al) inference, or Software as a Service (SaaS) applications on the cache nodes.

7. A method in a decentralized Internet Service Provider (ISP) access layer-based Content Delivery Network (CDN), comprising:Attorney Docket No.: 125512- 297358identifying an optimal subset of a cache nodes of a plurality of low-power, low-cost cache nodes, based on locality, congestion and performance history, wherein the plurality of low-power, low-cost cache nodes are deployed at or near end-user premises within an access network, and the cache nodes are located in at least one of residential units, optical splitters, or at a headend of the access network; andat a client device, requesting content and concurrently retrieving content segments from the subset of cache nodes using a receiver-driven coordination protocol, wherein the coordination protocol schedules segment downloads in forward and reverse directions to aggregate throughput and minimize head-of-line blocking.

8. The method of claim 7, wherein the cache nodes operate at a low power consumption of less than 15-30 watts.

9. The method of claim 7, wherein the coordination protocol is a Multi-Path Reverse and Forward Direction (MPRD) protocol.

10. The method of claim 9, wherein the client determines a download direction for each cache node and adjusts segment offsets dynamically based on node performance.

11. The method of claim 7, wherein the cache nodes are hosted in incentivized residential homes.

12. The method of claim 7, further comprising deploying micro CDN or edge computing workloads including at least one of video streaming, Artificial Intelligence (Al) inference, or Software as a Service (SaaS) applications on the cache nodes.

13. A tangible, non-transitory computer-readable medium comprising instructions that, when executed by one or more processors, cause the one or more processors to perform the method of claim 7.

14. A method for access-layer delivery in a distributed Internet Service Provider (ISP) environment, comprising:identifying a plurality of local cache nodes that store segments of requested content item;Attorney Docket No.: 125512- 297358selecting a subset of the cache nodes based on congestion and reliability metrics; instructing a client device to download non-overlapping content segments from the subset using a combination of forward and reverse segment scheduling; andaggregating throughput from the cache nodes to deliver high-resolution content without backbone bandwidth expansion.

15. The method of claim 14, wherein the local cache nodes are deployed across at least one of fixed broadband, mobile 5G / 6G, or satellite access network.

16. The method of claim 14, wherein each cache node is managed and authenticated by at least one of the ISP or a content provider.

17. The method of claim 14, wherein the method is further used for workload orchestration of cloud services including at least one of Artificial Intelligence (Al) inference, video conferencing or Application Programming Interface (API) gateway functions.

18. A tangible, non-transitory computer-readable medium comprising instructions that, when executed by one or more processors, cause the one or more processors to perform the method of claim 14.

19. A content delivery system operating within a satellite spot beam, wherein multiple incentivized homes each host a cache node, and a client downloads content segments from the nodes using a Multi-Path Reverse and Forward Direction (MPRD) protocol coordinated by a beam-local workload orchestrator.