Developer Toolkit

DOCA SDK & Programming

Build custom network functions on BlueField DPUs with NVIDIA's Data Center Infrastructure On a Chip Architecture (DOCA) software development kit.

50+
Core APIs
C/C++
Primary Languages
12
Hardware Accelerators
10×
vs Software-Only
Architecture

DOCA Software Stack

DOCA provides a unified programming model for BlueField DPUs, abstracting hardware complexity while enabling direct accelerator access for maximum performance.

Application
Custom Apps
Your network functions
Reference Apps
Firewall, Gateway, Switch
↓
DOCA SDK
DOCA Flow
Packet processing
DOCA Crypto
Encryption/decryption
DOCA Compress
Data compression
↓
Drivers
DPDK
Fast packet I/O
RDMA
Low-latency transfers
↓
Hardware
BlueField DPU
ARM Cores + Accelerators + Network
APIs

Core DOCA Libraries

Each DOCA library provides specialized functionality for common network and data center operations.

🌊

DOCA Flow

High-performance packet steering and modification. Define flow rules that execute at wire-speed in hardware.

doca_flow_pipe_create() doca_flow_pipe_add_entry() doca_flow_query()
🔐

DOCA Crypto

Hardware-accelerated encryption and decryption. AES-GCM at line rate with minimal CPU overhead.

doca_aes_gcm_encrypt() doca_sha_hash() doca_ec_sign()
📦

DOCA Compress

Hardware compression for storage and network. LZ4/Zstd at 100+ GB/s with dedicated engines.

doca_compress_deflate() doca_decompress() doca_compress_lz4()
🔍

DOCA RegEx

Pattern matching at wire-speed. Deep packet inspection, IDS/IPS, and application identification.

doca_regex_compile() doca_regex_search() doca_regex_match()
⚡

DOCA DMA

High-speed memory transfers between host and DPU. Zero-copy data movement with RDMA semantics.

doca_dma_memcpy() doca_dma_host_to_dpu() doca_dma_async()
📊

DOCA Telemetry

Real-time performance monitoring. Hardware counters, flow statistics, and custom metrics export.

doca_telemetry_start() doca_telemetry_export() doca_counter_read()
Example

Flow Rule Programming

Creating a hardware-offloaded flow rule for tenant isolation with QoS enforcement.

tenant_isolation.c C
1 #include <doca_flow.h>
2 #include <doca_flow_net.h>
3
4 int create_tenant_isolation_rule(uint16_t tenant_id, uint64_t rate_limit) {
5 struct doca_flow_match match = {
6 .outer.l2_type = DOCA_FLOW_L2_TYPE_ETH,
7 .outer.vlan_id = tenant_id, // Match tenant VLAN
8 };
9
10 struct doca_flow_actions actions = {
11 .meter.cir = rate_limit, // Committed rate (bps)
12 .meter.cbs = 12500000, // 12.5MB burst size
13 };
14
15 return doca_flow_pipe_add_entry(pipe, &match, &actions, NULL);
16 }
Models

Programming Models

DOCA supports multiple programming models to match your performance and complexity requirements.

Recommended

Pipe-Based (DOCA Flow)

Define flow rules as pipes with match criteria and actions. Hardware automatically processes matching packets at wire-speed without CPU intervention.

Latency
<5μs
CPU Usage
0%
Throughput
400 Gbps
Complexity
Low
Advanced

Run-to-Completion (DPDK)

Full control over packet processing. Poll packets from queues, process in software, and transmit. Maximum flexibility for custom logic.

Latency
10-50μs
CPU Usage
100%
Throughput
100 Gbps
Complexity
High
Pipeline

Packet Processing Pipeline

How packets flow through the DOCA-programmed BlueField DPU with hardware acceleration.

Hardware-Accelerated Packet Flow
📥
Ingress
<1μs
→
🔍
Parse
1-2μs
→
🎯
Match
1-2μs
→
⚡
Action
1-3μs
→
📤
Egress
<1μs