Whitepaper Exploring Computing Platforms for Radio Access Networks Heterogeneous Computing at the Network Edge Andy Butcher, Technical Staff, Server Advanced Engineering, Dell EMC Joseph Boccuzzi, Principal Wireless Access Architect, Network & Custom Logic Group, Intel Corporation Abstract A commentary is provided on server workloads pertinent to the wireless communications industry including 4G and 5G cellular.
Executive Summary Communications service providers are envisioning increased demand for mobile services including media and content delivery, mobile gaming, augmented and virtual reality, and connected vehicles. To satisfy this emerging demand, the buildout of 5G cellular infrastructure has commenced. Computing infrastructure is behind the scenes supporting radio access networks and other core services so this wireless ecosystem can grow.
Table of Contents 1 Introduction ......................................................................................................................... 4 2 C-RAN Background ............................................................................................................ 4 3 3GPP 4G Network Architecture and Software Stack ........................................................... 7 4 5 6 3.1 4G Network Architecture .....................................................................
1 Introduction A collective of communications service providers wrote a seminal whitepaper in 2012 describing a desire to transform their infrastructure to utilize standard IT equipment (servers, switches, storage appliances) to perform functions traditionally accomplished by customized equipment like routers, firewalls, and radio access network nodes.1 The concept of NFV, Network Functions Virtualization, was born.
Figure 1: Traditional Radio Access network (4G LTE) Referring to Figure 2, the following attributes can be associated with a Centralized Radio Access Network: 1. The protocol stack implemented by the BBU can be split in different ways between a Centralized Unit and a Distributed unit, with different implications and tradeoffs for bandwidth and latency. (Note the protocol can also be split in a three-tiered architecture including a CU, DU, and remote radio head RRH.
fewer handoff failures and less network control signaling, which can also be a potential savings with less need for inter-base station networks. As the concept of C-RAN evolved, the movement to virtualize workloads also pertained to this area, and the term vRAN (Virtualized Radio Access Networks) was created.
3 3GPP 4G Network Architecture and Software Stack 3.1 4G Network Architecture To do this workload study, it is necessary to describe the architectural elements and software stack in further detail before diving into performance optimization. In this section the LTE network (RAN + ePC) architecture elements are identified.
BBU Etc. FH RRH ... RF MIMO (BF) PDCP RLC DAC / ADC S1-MME BH MME S6a HSS MAC DFE, ET, DPD PHY S1-U S-GW S5 P-GW SGi Gx PCRF ePC eNB Figure 4: LTE Network Architecture block diagram The FH bandwidth utilization has been increasing with the cell capacity. The required bandwidth, now and in the future, brings forth complexity and costly solutions. Figure 5 illustrates network element partitioning options for the RAN to help mitigate this increase in FH capacity requirements.
3.2 4G Software Stack To continue with developing the context for the workload optimization study, this section discusses how the software components are distributed across the 4G network elements. First, Figure 6 depicts the software components for the control plane. An example of a Mobility Management (MM) transaction flow is provided between the UE (user endpoint) and the ePC. The interface between the eNB and MME is defined by the S1-MME protocol (control plane).
3.3 4G Edge Services Network Architecture The trend with 4G has been to adopt the cloud concept by centralizing compute and storage resources deep in the network. This approach works well when applications are not demanding higher data rates and lower latency. To address the overall end-to-end latency, 3GPP Release 14 introduced the concept of Control and User Plan Separation (CUPS) to allow a more flexible and scalable network.
MME S6a S11 S1-MME S5-C BBU ... RF MIMO (BF) DAC / ADC P-GW-C Sxa Sxb DFE, ET, DPD P-GW-U BH RLC MAC PHY PCRF Gx S-GW-C Etc. PDCP FH RRH ePC HSS MEC SGi S5-U Sxa SGi S1 S-GW-U S1-U P-GW-U S5-U MEC Edge Site eNB Figure 9: 4G Network Architecture Utilizing Edge Services. The well-defined interfaces between the eNB and ePC should not be barriers to collocating Edge functionality needed to satisfy service requirements.
techniques such as (ET, DPD), and Beamforming. The connection between the RRH and BBU is again called the FH. The FH interface is evolving from CPRI to eCPRI, IEEE1914.3, as well as O-RAN. The connection to the 5G CN is called the BH. The 5G CN will contain the Access and Mobility Function (AMF) which will perform Mobility Management, Access Authentication/Authorization, Terminate N2 interface, and Connection Management.
physical layer (PHY). The right side is the functionality represents the RRH functionality. The ORAN Alliance front haul specification defines User Plane, Control Plane, Synchronization Plane, and Management Plane multiplexing to allow for flexible and scalable access over the FH. DU/CU PDSCH Scramble Modulate Layer Map RRH Pre Code RE Map IQ Comp IQ Decomp Pre Code + BF iFFT + CP DAC ABF IQ Comp BF FFT - CP ADC ABF ... FH PUSCH Scramble Demodulate Chan Est RE Demap IQ Decomp ...
❖ The required per Cell FH BW = 15.3Gbps (w/o IQ compression or 9.2Gbps w/ IQ compression enabled). The required BH BW = 6.7Gbps (aggregated over 3 Cells). These configurations and results are illustrated in Figure 14: a) for LTE and b) for the 5G example. The respective FH rates have increased proportional to the user data rates.
Uu N2 UE AMF NAS Signaling MM MM RRC SDAP RRC SDAP N2-C N2-C PDCP PDCP RLC MAC PHY SCTP IP L2 L1 SCTP IP L2 L1 RLC MAC PHY gNB Air Interface Control Plane Figure 16: 5G NR control plane functionality The user plane software components are distributed across the network elements as shown in Figure 17. An example of a data transaction flow is provided.
5G CN UDM/PCF N2 AMF N11 SMF BBU MEC Etc. SDAP FH RRH (AAS) ... RF MIMO (BF) DAC / ADC DFE, ET, DPD UPF N6 PDCP RLC MAC BH N4 PHY N3 N9 UPF Edge Site N6 MEC gNB Figure 18: 5G Network Architecture Using Edge Services. Edge computing benefits users when the time delay of transporting packets is significant. When compute functionality is placed closer to the edge, there is significant reduction in latency.
Figure 19: 5G Network Highlighting Network Slicing Example. 4.6 Software Offload Functionality In viewing the functionality required in the RAN (CU & DU), various functions lend themselves to hardware acceleration. The best candidates are in the PHY layer and include forward error correction (FEC) and I/Q sample compression. The FEC functionality can be accelerated to free up compute resources for other functions.
FEC Encoding o Code Block CRC Generation, Turbo Encoding (4G), LDPC Encoding (5G) and Rate Matching (Sub-block interleaving, Bit Selection/Collection). FEC Decoding o HARQ Combining, Rate De-Matching (Sub-block de-interleaving, Turbo Decoder (4G), LDPC decoder (5G) and Code Block CRC check.
Figure 21: Test setup in Dell lab, with Layer 1 software modules 5.2 Results The benefit of offloading the Turbo operations to an FPGA was explored. The device chosen for the offload was a prototype version of the N3000 network card13 because it allowed a unique combination of “inline” and “lookaside” operations. Turbo FEC, because of its position in the stack, is not suited for an inline offload.
be seen in Table 1, which shows normalized cycles attributable to the uplink decode and the downlink encode. As with any real-time computing task, another important timing consideration is the deviation in processing time. In datacenter applications, this has been called the “long tail” in latency. This test setup was not equipped to measure the standard deviation, however the maximum duration was captured for the thread that performs the interleaving portion of the Turbo algorithm.
6 Platform Architectures 6.1 Flexible Architectures for the Network Edge Using the 5G NR network architecture split, Figure 23 can be used to conceptualize server platforms that utilize FPGA peripherals. In the DU area, the FPGA will be used, as demonstrated, for look aside acceleration such as 4G/5G FEC.
Figure 24: Intel® Xeon® Scalable processor-based solution This solution is available with a single mother board design where the PCIe card functions can be reconfigured depending on this platform location and function in the network. For use in DU, CU and Edge Cloud deployments, the following examples are given. Depending on space availability, various platform configuration optimization can be used.
6.3 Products 6.3.1 Rack Servers Mainstream servers in the PowerEdge portfolio from Dell EMC can be used to implement radio access networks, and as mentioned as one of the key tenets of NFV workloads, the economy of scale that is achievable by using off-the-shelf equipment is attractive. The R640 has been validated by the Dell OEM team for NEBS Level-3 compliance.17 It is a 1U two-socket rack server, offering a high degree of computing density.
9 References 1 https://portal.etsi.org/nfv/nfv_white_paper.pdf 2 https://doc.dpdk.org/guides/prog_guide/bbdev.html 3 www.3gpp.org China Mobile paper 4 https://pdfs.semanticscholar.org/eaa3/ca62c9d5653e4f2318aed9ddb8992a505d3c.pdf 5 https://www.o-ran.org/ 6 http://www.cpri.info/index.html 7 https://standards.ieee.org/standard/1914_3-2018.html 8 3GPP TS29.244 (v15.5.0), Interface between the Control Plane and User Plane Nodes, www.3gpp.org 9 3GPP TS23.501 (v16.0.