Does the cloud-native 5G core network need a DPU?

In 2022, the number of 5G connections worldwide will exceed the 1 billion mark, and each 5G user will consume twice as much data as a non-5G user. Mobile network traffic is growing rapidly. Analysis agency Omdia predicts that by 2023, 5G users will consume an average of 14GB per month, and by 2027 this will double to 28GB.

Operators need to consider how to handle this rapidly growing traffic.

5G core network evolves towards cloud native

At present, the design of 5G core network is based on NFV virtualization architecture, which consists of VNF (virtual network function). VNF is a software package built on general hardware and is a network function running in a virtual environment.

The concept of NFV was first proposed by the ETSI organization. It refers to the use of IT virtualization technology to realize various network equipment functions using standardized general-purpose IT equipment. Its essence is to achieve the decoupling of hardware resources and software functions. The goal is to replace private and dedicated network element equipment in the communication network with standard x86 servers, storage and switching equipment, saving investment for operators while realizing the rapid development and deployment of new services, and realizing automatic deployment, elastic scaling of capacity, fault isolation and self-healing based on actual business needs.

As shown in the figure, two PNFs, Firewall and VPN Gateway, need to use two virtual machines to run this VNF. In addition, the hypervisor of the virtual machine divides the physical server into multiple logical servers (VMs). At the same time, the virtual machine also needs to run the Guest OS operating system for each virtual machine on the host machine, which is an additional burden on the server.

However, VNFs that are separated from dedicated hardware are large-grained telecommunications software packages that are very large and complex, often involving millions of lines of software code. This means that the entire process from software development to release and testing is extremely labor-intensive and is estimated to take a year.

In order to meet the requirements of large bandwidth and low latency in 5G application scenarios, the user plane of the 5G core network must not only be deployed to the edge or regional data center to reduce transmission latency, but also significantly reduce the forwarding latency of user-plane messages. VNF (virtual network function) can no longer meet the demand quickly.

What should we do? Based on the design principles of cloud native, we can further decompose the large-grained VNF into multiple small-grained microservices. Microservices are not only small-grained, but also have independent lifecycle management, which can achieve more fine-grained software development, release, testing and upgrade, thus improving operational agility.

Cloud Native defines a new way to develop and run applications in a virtual cloud environment. Cloud Native principles include the following:

• Applications are "split" into smaller units called microservices. A set of smaller, interconnected microservices replaces a single application.
• Containers hold microservices and provide a runtime environment. Containers package application code, binaries, and dependencies without the overhead of a virtual machine. Containers share the same Guest OS or Host OS system/kernel.
• Kubernetes orchestration provides complete container lifecycle management, including scheduling, start/stop/restart, and visibility.

CNF also has other outstanding advantages, including:

• Improve flexibility and agility, as launching new services or upgrades no longer involves replacing any hardware.
• Lower costs, as program deployment requires less hardware than what is needed to support VNFs.
• Improved scalability, because containerized microservices can be scaled horizontally or vertically as needed. Due to the nature of the cloud, it is easy to deploy more hardware resources to support large amounts of traffic or concurrent users, which can basically be considered to have unlimited scalability.
• Improve fault tolerance and reliability. If a container goes offline for any reason, operators can immediately start another container.

Improving Network Performance: Software or Hardware?

The industry has long debated how infrastructure originally designed for computing tasks can effectively handle 5G user plane and low-latency edge use cases?

One of the key discussions is: use hardware accelerators (aka SmartNICs or data processing units (DPUs)) to offload packet processing to more efficiently utilize server resources, or look to software to optimize performance?

There is a view that software is still likely to be the primary way to gain efficiency, namely through carefully designed cloud-native network functions (CNFs), and that this is the better approach in the near term because it retains the agility and workload portability that can make the cloud more powerful.

The hardware solution refers to reducing the CPU load by distributing processing work to acceleration hardware (coprocessor) and utilizing the processing advantages of suitable acceleration hardware to achieve performance improvement and cost optimization.

The hardware under NFV generally uses x86 general-purpose servers, but the I/O performance of x86 general-purpose servers, such as throughput and latency, is far inferior to that of traditional dedicated hardware and cannot meet the commercial deployment requirements of 5G scenarios. On the surface, hardware acceleration is attractive in the 5G core.

Using external hardware such as DPU/SmartNIC to offload data packet processing can improve flow table lookup and message processing capabilities. For data flow messages that need to be accelerated, there is no need to upload them to the CPU. They can be processed and forwarded locally on the DPU/SmartNIC, achieving hardware-level processing and forwarding of data messages, minimizing the use of business messages. This can greatly reduce the use of CPU resources, significantly improve performance and reduce latency.

In addition, at the virtualization layer, the decoupling of the business layer, virtualization layer, and hardware layer of the existing NFV architecture is progressing slowly. We can explore sinking the virtualization layer to the DPU to make the NFV architecture more open.

Currently, operators are promoting open UPF, but a large number of UPFs are currently bound to equipment manufacturers. The use of DPU/SmartNIC can realize the offloading of UPF's GTP, QoS, Forward and other functions, improve UPF forwarding efficiency, increase single server throughput, reduce processing delay, jitter and packet loss rate, etc., and can reduce the number of user plane servers deployed at each site, thereby reducing the requirements for room space, power consumption, and heat dissipation, and reducing costs.

DPU/SmartNIC for 5G Core Network

The following figure shows the user survey results on "Does your organization expect to make extensive use of hardware acceleration (DPU/SmartNIC) in its 5G core user plane?"

40% of respondents chose “Use DPU/SmartNIC in specific scenarios such as fixed wireless access”. Fixed access services generate much higher throughput per connection than mobile services (about 10 times higher as a rule of thumb). Therefore, the case for using hardware acceleration is stronger. Although it will take a few years, more people will choose to use hardware acceleration if wired and wireless begin to converge and a single user plane is deployed for fixed and mobile access.

38% of respondents selected “Extensive use of DPU/SmartNIC, user plane acceleration is essential in most cases”, predicting that hardware acceleration will be important for efficient processing of 5G core user plane traffic.

Swinging

DPU hardware acceleration has many advantages, the most significant of which is that it can greatly improve network performance and provide higher speeds, lower latency, and stronger network performance than traditional technologies. This will enable 5G networks to transmit data faster and more reliably.

Although DPU/SmartNIC hardware acceleration has certain advantages in 5G core networks, the industry was once skeptical about whether the user plane of the core network NFV needs to use hardware acceleration.

Virtualization of general hardware platforms is an inevitable trend. Operators have just transitioned away from dedicated equipment, and DPU/SmartNIC seems to have returned to familiar dedicated hardware.

Hardware acceleration performance ratio does not meet expectations: The main reason why dedicated equipment is criticized is that the monopoly of equipment vendors has caused high costs. At present, the performance improvement and cost reduction of core network hardware acceleration are limited. At the same time, the hardware binding problem that may be caused by the introduction of acceleration hardware makes it difficult for operators to make up their minds.

There are also rumors that the current 5G traffic load does not really require the deployment of DPU/SmartNIC acceleration technology. With good software design and wise deployment choices, 5G mobile user plane traffic can be easily handled on standard server hardware.

Of course, there are also views that the coordinated acceleration of software and hardware is the general trend.

If only a few places/services currently require hardware acceleration, operators must comprehensively consider whether they need to invest in this technology now. Therefore, in general, operators are interested in hardware acceleration of the user plane of the 5G core network, but the current situation is not clear.