NEC's UPF maximizes

5G value with high performance

and flexibility in containerized,
virtualized or physical deployments

NEC UPF benefits that Key features of NEC UPF

realize the 5G potential • 3GPP compliant
　

• 5G/4G/3G converged nodes

5G core can be decomposed into control and user
　

(UPF+PGW-U/SGW-U+GGSN-U)
planes. The control plane (C-plane) nodes provide • IPv4, IPv6 and IPv4/v6 packet
　

communication control functions while User Plane • I-UPF and MEC

　

Function (UPF) handle user data traffic. In 5G • Offline/online billing processing

　

architecture, the C-plane and U-plane, which have • Rate limit

　

traditionally been collocated, are now completely • QoS policy definition for IP traffic based on DSCP
　

decoupled. This is known as Control and User Plane • N+1 node redundancy
　

(session continuity and rapid switching)

Separation (CUPS).
• Built-in basic DPI functions with advanced
　

CUPS enables control and user planes to be DPI plug-ins

independently scaled and deployed. This allows, • Multi-platform support
　

for instance C-plane nodes to be located on a (virtualized/containerized environments)

centralized cloud, so that only the UPF and Radio • Network Slicing
　

Access Network (RAN) are deployed closer to the

edge of 5G coverage areas, reducing latency and designed in a cloud-native architecture to meet
increasing capacity. these requirements. NEC’s strong track record and
As a result, service providers can offer C-plane technical capabilities and expertise in both mobile
functions as managed services, which will networking and computing domains, enables a 5G
significantly reduce the load of installation and product strategy that focuses on market needs.
operations when deploying additional 5G coverage
areas.
Market-leading performance
Having said that, UPF needs to meet several In June 2021, NEC successfully conducted a
requirements to bring out the true potential of 5G. measurement test, which demonstrated that its
For example, 5G is expected to be several times UPF could handle traffic at the rate of 640 Gbps
faster than 4G LTE, with a peak speed of up to 20 per server. NEC verified this by using 2U/2-socket
Gbps downlink per device, along with high-reliability servers equipped with the latest server processors
communication and ultra-low latency of less than and 100 GbE Network Interface Cards (NIC). The
1 milli second. High-performance UPF is crucial to UPF capability demonstrated in this measurement
realize this promise. test was significantly better than traditional systems.
In addition to this, 5G also needs to support different Moreover, NEC UPF has the flexibility to support
deployment models. It should be possible to deploy multiple platforms. It can run on virtualized or
UPF on different scales and platforms to support not containerized environment, as well as on a COTS
only commercial 5G communication services offered server. It can also be deployed on public cloud,
by mobile network operators, but also industrial which enables the quick launch of 5G network on
solutions that use local 5G or private 5G. a small scale with short-lived use. NEC UPF is a
The UPF developed and delivered by NEC has been converged solution of 4G/3G U-plane functions

1
(PGW-U/SGW-U+GGSN-U) on top of 5G U-plane
function, enabling investment cost reduction.
This whitepaper introduces the details of NEC’s
high performance UPF and the technology behind
together with our extensive product lineup and the
use cases.

2
Performance test results
for containerized UPF software

NEC’s containerized UPF software (8 pods per

Industry-leading high performance server) runs in an environment that consists of a
verified for NEC UPF general-purpose x86 server with 3rd Gen Intel®
NEC has developed and deployed a commercial Xeon® Scalable processors (32 cores x 2 sockets)
containerized UPF software and demonstrated and four dual-port 100 GbE NICs, along with a
industry-leading traffic handling at 640 Gbps per platform configured on Linux and Kubernetes. UPF’s
server (320 Gbps per CPU) by using that. The detailed processing capability was measured by using test
conditions of the measurement environment and traffic that simulated the typical commercial traffic of
test results are described below. the mobile network services.

Measurement environment Measurement conditions

Intel® Xeon® Throughput (Gbps) 640 Gbps
CPU Platinum 8358 processor
2.60 GHz, 32 cores×2 sockets No. of sessions 400K sessions

Memory DDR4-2933 DIMM (8ch/CPU) No. of FARs 2 FARs per session

NIC
Intel® Ethernet Network Adapter No. of QERs 2 QERs per session
E810-2CQDA2 ×4
No. of URRs 2 URRs per session
Host OS CentOS Linux 8.2
Ratio of UL/DL UL:DL = 1:3
Host OS kernel 4.18.0-193.28.1.el8_2.x86_64
User Packet Size 800 bytes
Kubernetes Ver. 1.19.3

DPDK Ver. 20.08

640Gbps/server

UPF Pod UPF Pod UPF Pod UPF Pod UPF Pod UPF Pod UPF Pod UPF Pod
80Gbps 80Gbps 80Gbps 80Gbps 80Gbps 80Gbps 80Gbps 80Gbps

200G NIC 200G NIC 200G NIC 200G NIC

(with DDP function) (with DDP function) (with DDP function) (with DDP function)

100G 100G 100G 100G 100G 100G 100G 100G

3
NEC’s high performance
UPF architecture

These CPU cores process instructions sequentially

Industry-leading performance with from multiple threads based on OS scheduling.
NEC’s expertise and technical skills However, if a UPF application is added here, tasks
Now, let us explain how NEC has achieved such a may be interrupted by other threads, which can
high performance of UPF. sometimes cause packet forwarding process to wait.
NEC uses Intel’s technologies of Data Plane NEC uses Linux’s CPU Pinning function in UPF packet
Development Kit (DPDK) and Dynamic Device processing threads to pin (bind) them to CPU cores.
Personalization (DDP). NEC performed significant In other words, specific CPU cores are exclusively
optimization and improvements to be able to get dedicated to a packet processing purpose only, in
the highest possible performance for the cloud order to improve the efficiency and performance.
native UPF. The six key aspects of our approach are Moreover, NEC is using DPDK’s NIC Poll Mode Driver
described below. (PMD). Usually, the OS reactively begins processing
only after the NIC gets a notification about received
Maximizing CPU utilization with CPU packets. PMD improves this by proactive and
1
Pinning and Poll Mode Driver constant polling to process received packets quickly,
A normal Linux system runs multiple applications leading to stable and high-speed packet forwarding
(processes and threads) on a shared CPU core. by UPF.

UPF software
Core
Core #N
#1
Core
GTP PDR QoS Charging
GTP #0
Decap LookUp (Policing/Marking) Encap

Maximizing CPU
utilization CPU socket #0 CPU socket #1
Main Main
memory memory

Core Core Core Core Avoiding communication between CPU sockets

Faster memory
access (HugePages)
NIC
Faster memory
access (prefetching) DDP function Offloading to NIC the user identification processing

Optimizing CPU
processing through
bulk processing

Packet flow

4
 processing.
Faster memory access
2 Packet forwarding in UPF essentially repeats the
with HugePages
same processing steps for each packet. NEC decided
Linux provides a virtual memory space that abstracts to aggregate multiple packets at the same time in
the physical memory space, making it easier for each processing step so that they are processed
applications to handle. Virtual and physical memory in bulk. This boosts the efficient use of instruction
spaces are divided into pages (indicating size), and caching and improves the throughput accordingly.
each page is managed by assigning an address
number to it. The default size of a page is 4 KB. Avoiding communication
5
When an application needs to look up data on the
between CPU sockets
memory, it specifies the data location by using the In a server with two Intel® Xeon® Scalable
virtual memory address. Then, the processor looks processors, the memory is split and connected under
up the conversion table on the kernel to translate the each CPU socket. So when the memory connected
virtual memory address into the physical memory under processor B is accessed from processor A,
address and accesses the data after understanding communication takes place via Intel® Ultra Path
its location on the physical memory. Interconnect (UPI), which is a communication path
When the application is handling a huge volume that connects processors. However, UPI capacity
of data, the frequency of address translation by the can be a bottleneck in processes that handle large
kernel becomes high if the page size is small, which volumes of data, slowing down the processing
can degrade the application performance. UPF is speed. To overcome this bottleneck, NEC’s UPF
one such application that consumes a huge volume software is designed to prevent memory access
of data, with gigabytes of memory to perform a large across UPI.
amount of packet forwarding.
To improve the situation, NEC has enabled Linux’s
Offloading to NIC the user
6
identification processing
HugePages function and configured the UPF
memory’s page size to 1 GB. This allows a huge To increase the efficiency of packet forwarding in
memory to be looked up efficiently, reducing the UPF, the same user’s packets must be sent to the
processing overhead. same CPU core for user-specific processing.
To achieve this, until now the user identification
Faster memory access for received packets was performed in the UPF
3
by prefetching data
software, which was then distributed to the same
Prefetching is a process in which the data that a CPU core that is assigned to handle a specific user’s
program references or uses is fetched from the traffic. But in this case, a switch would occur between
main memory (DRAM) to the CPU’s cache memory the core that performs first-level of processes for
(D-Cache) ahead of time. If the program fetches the user identification, and the core that performs
data by accessing the main memory after a process subsequent user-specific processes of the packet
starts, that process will be delayed by the time taken forwarding. The core was poorly utilized, and the
to access DRAM. Prefetching can reduce this delay. process itself was complicated.
NEC’s UPF proactively uses prefetching to reduce Now, NEC has offloaded the user-identification
processing overhead, which boosts the processing process to NIC by using the Dynamic Device
speed. Personalization (DDP) function provided by Intel NIC
(E810). The DDP function performs user identification
Optimizing CPU processing
4 for a packet on the NIC, and forwards the packet to a
through bulk processing
specific core assigned to that user.
The CPU has a mechanism to cache application The DDP function improves the utilization of CPU
instructions (I-Cache) just like data cache. If a process cores because the UPF software doesn’t need to
needs to be executed multiple times, instructions perform the user identification process anymore. In
are cached the second time onwards for faster addition, packets no longer need to be redistributed
processing. Processing instructions multiple times (switched) between cores, reducing the time lost
simultaneously with this mechanism is called bulk in the switching process. The overall software

5
processing has become simpler, with significant
improvements in processing efficiency and speed.

Traditional architecture
Rx (received) task Worker task
Core#1 thread Core#N+1 thread
Rx Rx User Packet Soft User-specific Tx Tx
NIC queue 1 process identification redistribution queue process process queue 1 NIC
Core#2 thread Core#N+2 thread
Rx Rx User Packet Soft User-specific Tx Tx
queue 2 process identification redistribution queue process process queue 2
RSS Core#N thread Core#N+M thread
Rx Rx User Packet Soft User-specific Tx Tx
queue N process identification redistribution queue process process queue M

First level of packet distribution: Distribute packets Second level of packet distribution:
to each core without identifying the user Re-distribute packets based on user identification

Optimized architecture
Rx (received) + Worker task
Core#1 thread
Rx Rx User-specific Tx Tx
NIC queue 1 process process process queue 1 NIC
Core#2 thread
RSS
Rx Rx User-specific Tx Tx
queue 2 process process process queue 2

DDP Core#L thread

user iden- Rx Rx User-specific Tx Tx
tification queue L process process process queue L

Distribute packets to each core based on Offload to NIC the user Packets don’ t need to be re-distributed between cores,
NIC’s (DDP function) user identification identification processing improving processing efficiency and maximizing CPU utilization

6
UPF product lineup for a variety of
deployment scenarios and use cases

1BOX-UPF, a COTS server consisting of UPF software

Support for container, virtual, installed on an x86 server, uses custom tuning to
and physical platforms extract the best hardware performance, delivering
NEC's UPF is designed to be cloud native and high throughput and session-processing capabilities
hardware agnostic. It can be deployed on a wide in a 2U server. Our product lineup also includes
range of platforms. a small UPF, a compact server that is less than
Either containerized or in virtualized environments half the size of 1BOX-UPF. The small UPF can be
(i.e. NFV: Network Functions Virtualization), the easily installed in environments such as local 5G or
UPF software can be deployed in CSPs data temporary setups for one-time events.
centers, either centrally or at the edge, but also
in enterprise data centers as well as in the public
5G/4G/3G Integrated U-plane
cloud. Leveraging the benefits of cloud native UPF The converged UPF integrates 4G/3G U-plane
allows flexible and rapid deployment scenarios processing functions (PGW-U/SGW-U+GGSN-U), in
based on traffic and network characteristics, addition to 5G UPF functions. For example, while
including deployment on the same infrastructure migrating to a 5G network from a 4G/3G network,
as CU/DU/MEC. Moreover, its high performance processing of the mixed user traffic (i.e. 5G/4G/3G
enables extremely efficient use of power and server- traffic) can be easily handled by this converged UPF
installation space in data centers, helping CSPs to alone, preventing the duplication of investment and
reduce their costs. operating costs.

Multi-deployment model in NEC UPF

B2B service Enterprise local Edge model Central model

data center model

RAN UPF

RAN

B2C service UPF

RAN

5G core
RAN UPF (control plane)

7
 used in various scenarios – for example, to perform
Realizing B2B2X business models consolidated analysis of a factory’s operation and
NEC UPF easily supports a wide range of use cases management by connecting to the company’s data
because it can be deployed on a variety of platforms. network via UPF.
In addition to the commercial networks of mobile Further, the U-plane can be flexibly deployed
network operations and local 5G of enterprises, it is wherever needed in the network, so cost-effective
also expected to make business development easier and optimal U-plane deployment is possible in
for the B2B2X model in which CSPs collaborate with usage scenarios with different requirements like high
industry partners to deliver solutions for various bandwidth and ultra-low latency.
industries.
For example, if a 5G solution for smart factories is to
be delivered through a B2X model, C-plane can be In addition to UPF, NEC’s 5G core network consists
offered as a managed service by deploying the 5G of containerized cloud-native components, and
core on a central data center, while the UPF and RAN uses 3GPP-compliant open architecture. 5G
equipment are deployed in the user environment is a technology that is likely to be deployed in
(factory). diverse businesses. In this context, NEC’s 5G
In this case, U-plane traffic stays within the local products facilitate rapid, flexible, and powerful 5G
network, so that you can securely handle data that deployments for users in all scenarios.
you don’t want to share externally. This can be

The names of products and services in this document are trademarks

or registered trademarks of their respective owners.

Published by Network Services Business Unit, NEC Corporation

https://www.nec.com/en/global/solutions/5g/
June 2021

© NEC Corporation 2021