In-network Service Acceleration Platform (ISAP)

2.1 Event-driven resource deployment

Various applications are expected to use the ISAP, and each application requires a wide variety of resources and processing functions. It is inefficient to deploy all processing functions in the network in a fixed manner. Therefore, the ISAP uses a method that allocates the necessary amount of resources only when users are running applications. Specifically, the ISAP collects various information such as terminal location registration, communication sessions, control events such as handover, application authentication and service-startup events, and activities in cyberspace. It then analyzes this information and allocates the required amount of computing resources to applications at the required time and location. This method enables efficient use of resources for various applications and processes.

2.2 Inter-accelerator chaining using dedicated hardware

Applications envisioned with the ISAP, such as connected cars, robotics, VR, and telesurgery, involve advanced computational processing, including rendering, artificial intelligence (AI) image analysis, and cryptographic computation. The results of the processing must be transmitted to the terminal via a network. Therefore, both application processing and transfer must be completed quickly and with low latency. The ISAP uses dedicated accelerators such as GPUs, field-programmable gate arrays (FPGAs), and data processing units (DPUs)/smart network interface cards (SmartNICs). More specifically, the ISAP creates microservices for various application-processing and network-connectivity functions and assigns an appropriate accelerator to each. A GPU, for example, may be allocated for AI analysis and three-dimensional (3D) rendering processing functions, and a DPU/SmartNIC for GPRS (general packet radio service) Tunnelling Protocol (GTP) encap-decap and Real-time Transfer Protocol (RTP) streaming transmission/reception functions. These applications and network functions in the form of microservices are chained together by hardware-accelerator chaining to achieve central processing unit (CPU)-independent processing. Therefore, the ISAP enables high-speed and low-latency processing of applications and transfer of processing results within the network. Remote cooperative motion control can also be accurately and precisely achieved.

2.3 Fault detection and visualization for system resilience

The convergence of networking and computing brings about an increase in overall system complexity. The 3rd Generation Partnership Project (3GPP) has also raised the issue of resilience of core networks in mobile communication networks, and future countermeasures are being discussed. The ISAP has a mechanism to detect and quickly mitigate failures and anomalies such as signaling storms in the network. By visualizing the location of failures and the causes of anomalies and detecting signs of failure, the ISAP can prevent the spread of signal congestion or even prevent signal congestion, enabling a robust network system. For more detailed information, please refer to the article on robust networks [3].

3.1 Implementation

To verify the feasibility and effectiveness of the ISAP, a prototype implementation was conducted along with use cases. The system configuration is shown in Fig. 3. The two use cases were AI-based video-stream analysis and metaverse.

Fig. 3. Implementation configuration.

Feature (1): event-driven resource deployment was achieved by connecting the ISAP to the 5G core network (5GC) and applications in the cloud. More specifically, the ISAP is connected to the 5GC using the 3GPP standard interface [4] to collect information on user terminal activation, network registration, and movement. The ISAP is also connected to the application to collect information such as application startup and state transition. On the basis of the collected information, the ISAP detects application startup and deploys AI analysis and video rendering functions to the network on demand.

Feature (2): inter-accelerator chaining using dedicated hardware is achieved using device plug-ins and custom resource functions. Various network functions, such as GTP header processing and RTP streaming, are containerized, and appropriate accelerators are virtualized and allocated to each container. High-speed processing without CPU intervention is also achieved by chaining between containers.

Through the above prototyping, we demonstrated the feasibility of flexible control and allocation of hardware resources in conjunction with user and application events.

3.2 Evaluation

(1) Effects of event-driven resource deployment

The effectiveness of event-driven resource deployment, the first feature of the ISAP, was examined. We verified the total memory usage when a large number of users randomly use various applications at the same time. The resource design required by each application was based on the actual container-functionality resource-sizing design. The experiments confirmed that the ISAP uses computing resources more efficiently than fixed memory-allocation methods. This is due to the ISAP’s feature of allocating computing resources to individual users when they need them.

(2) Effects of inter-accelerator chaining using dedicated hardware

Next, we verified the effect of offloading the application-processing and network-connection functions onto the accelerator and chaining between each function. The application processing performance was measured for the AI video analysis system shown in Fig. 3, and the network connection performance was measured for the user plane function (UPF) and tunneling processing in 3D video rendering shown in Fig. 3. The measurement results indicate a 97% improvement in latency and 99% improvement in jitter for packet transmission performance. For data processing performance, latency and jitter were reduced by 28 and 95%, respectively. It was also confirmed that the system is capable of high-performance analysis and smooth playback and operation of 3D images compared with conventional systems.

4.1 AI video analysis

As represented by the recent evolution of generative AI, AI technology is expected to be key to solving various social issues. For service providers to offer solutions incorporating their own AI model to users, however, it is generally necessary to build a framework that includes data collection and preprocessing, model learning and evaluation, and model deployment and operation. In particular, training and deploying a model with many parameters may require an AI cluster consisting of many GPUs, which can be very costly.

We have demonstrated an AI solution service that usually requires an advanced computing platform by using GPU and DPU resources managed by the ISAP. In this experiment, the low-latency and low-jitter capability of data processing leveraging the DPU and GPU, as described in Section 3, enabled very smooth and accurate AI video analysis even in use cases that require a high data framerate, such as 4K uncompressed video (Fig. 4(a)). By efficiently using the GPU and DPU resources deployed on the site of telecom operators, the ISAP can quickly and flexibly deploy AI models to meet the requirements of various service providers such as for mission-critical services.

4.2 Metaverse

The development of 3D computer graphics and mobile communication technology and the spread of remote work have increased the opportunities for users to communicate with each other online. Metaverse platforms that can accommodate many users have emerged and are expected to continue to develop and spread further.

We conducted a demonstration that assumes a future use case in which users come and go between multiple metaverse spaces [5]. Specifically, we confirmed the ability to dynamically and rapidly draw the destinated 3D virtual space on the user’s terminal as the user transitions in conjunction with the community world of MetaMe^TM [6], which allows users to move to various metaverse spaces (Fig. 4(b)). By using the ISAP to link the terminal-side network information (e.g., connection status to 5G access) and the cloud-side metaverse environment information provided by the service provider, it is possible to create a flexible service experience independent of terminal specifications, type of network access, and service environments.

4.3 Network slicing

Network slicing, in which virtual network resources are isolated from the shared physical infrastructure to satisfy various service requirements, is attracting attention as a method of monitoring and guaranteeing service-level agreements for telecom operators.

We demonstrated the provision of optimal end-to-end (E2E) network slices to meet the requested area, time, and purpose in cooperation with NTT DOCOMO’s 5GC deployed on a public cloud [7]. The ISAP’s event-driven resource control and hardware acceleration chaining based on the user network connection status provide high-quality E2E network services on demand, guaranteeing application and service quality for specific traffic congestion spots, such as event sites and disaster areas.

4.4 Fault detection and visualization for system resilience

To prevent the expansion and prolongation of failures in the 5GC, a mechanism that detects and visualizes failures occurring in the 5GC and feeds back to the 5GC for failure countermeasures and resource control will be effective. For this experiment, we re-designed the 5GC as a stateless application system so that the effects of failures do not spread to other applications or databases. We also constructed a demonstration environment consisting of a Stream Control Transmission Protocol (SCTP) termination unit, message delivery infrastructure (Kafka), control plane processor (5GC APL), and shared database (Fig. 4(c), (1)). In this configuration, even when 5GC APL is stopped, processing is executed in a shared manner. By acquiring information on the connection status of user terminals from the database, it was confirmed that processing could continue without problems.

Fig. 4. Demonstration experiments.

Assuming the occurrence of an unusual failure, we reproduced a situation in which the location registration process fails under conditions of extremely large packet loss and delay in only one direction of communication (Fig. 4(c), (3)). The user is unable to start communication because the location registration process is not completed. At the network layer, SCTP and Transmission Control Protocol (TCP) retransmissions occur, resulting in message retention and upper-layer retransmissions. We thus confirmed that understanding the behavior of messages and packets is effective in narrowing down the cause of the problem and determining the corrective action to be taken.