NVIDIA DPU Course

The NVIDIA BlueField Data Processing Unit (DPU) is a high-performance networking and compute accelerator designed for modern data centers. This post introduces the main components and features of the BlueField DPU.

Hardware platform

The NVIDIA BlueField DPU currently has the following main product generations:

BlueField-3 DPU
BlueField-2 DPU (available in InfiniBand/Ethernet dual versions)
BlueField-1 series controller cards

Software architecture

DOCA software framework

DOCA (Data center-on-chip Architecture) is the software framework developed by NVIDIA for the BlueField DPU, and includes:

DOCA SDK (v2.10.0)

RDMA acceleration SDK
Network acceleration SDK
Security acceleration SDK
Storage acceleration SDK
Data Path Accelerator (DPA) SDK
Management SDK

Main components

Unified Communication X (UCX)
RDMA verbs
GPUDirect
Software-defined networking (SDN)
P4 programming support
Inline encryption
App Shield runtime security

Application scenarios and features

1. Network acceleration

1.1 Network protocol stack offload

TCP/IP offload
- Complete TCP/IP protocol stack processing
- Supports TCP Offload Engine (TOE)
- Reduces host CPU load
- Reduces network latency
RDMA technology
- Supports RoCE v1/v2 (RDMA over Converged Ethernet)
- Supports InfiniBand
- Zero-copy data transfer
- Kernel bypass
- QP (Queue Pair) management
- Supports RDMA Write/Read/Send/Receive operations

1.2 Network virtualization

SR-IOV (Single Root I/O Virtualization)
- Supports up to 1000 VFs (Virtual Functions)
- Hardware-level network resource isolation
- Virtual machine passthrough
OVS (Open vSwitch) acceleration
- Hardware offload of flow table processing
- Supports the OpenFlow protocol
- Virtual switch performance optimization

1.3 Advanced networking features

Traffic management
- QoS (Quality of Service) support
- Bandwidth control
- Traffic shaping
- Congestion control
Network security
- IPSec hardware acceleration
- TLS/SSL offload
- Firewall rule processing
- DDoS protection

1.4 Network performance metrics

Bandwidth capability
- Supports 25/50/100/200/400GbE
- BlueField-3 supports up to 800Gb/s bandwidth
- Dual-port configuration options
Latency optimization
- End-to-end latency <1 microsecond
- Hardware timestamp support
- Precision Time Protocol (PTP) support

1.5 Network programming models

DPDK support
- User-space network drivers
- Poll Mode Driver (PMD)
- Zero-copy packet processing
- Huge page memory support
P4 programmability
- Custom packet processing pipelines
- Protocol-independent packet processing
- Flexible forwarding rule definition

Network layer model

应用层 (L7) - HTTP, FTP, SMTP等
传输层 (L4) - TCP, UDP
网络层 (L3) - IP, ICMP
数据链路层 (L2) - 以太网, MAC
物理层 (L1) - 物理介质

Key networking concepts
- MTU (Maximum Transmission Unit)
- VLAN (Virtual LAN)
- VXLAN (Virtual Extensible LAN)
- ARP (Address Resolution Protocol)
- BGP (Border Gateway Protocol)

1.7 Performance tuning suggestions

Network tuning
- Enable Jumbo Frames
- Configure RSS (Receive Side Scaling)
- Optimize interrupt affinity
- Use NUMA-aware memory allocation
Best practices
- Plan the network topology sensibly
- Choose the appropriate network mode
- Monitor network performance metrics
- Run performance tests regularly

2. Storage acceleration

NVMe storage virtualization
Storage encryption and compression
Zero-copy data transfer

3. Security features

Hardware-accelerated encryption
IPsec security gateway
Zero-trust security architecture
Runtime security protection

4. Virtualization support

SR-IOV virtualization
Virtual device emulation
Containerized workload support

Reference source
DOCA framework documentation

Translated from the Chinese original.