GPU Packet Processing

NVIDIA DOCA GPU Packet Processing Application Guide

This document provides a description of the GPU packet processing application to demonstrate using the DOCA GPUNetIO, DOCA Ethernet, and DOCA Flow libraries to implement a GPU traffic analyzer.

Note:

This application is currently supported at beta level.

Real-time GPU processing of network packets is a technique useful to several different application domains, including signal processing, network security, information gathering, and input reconstruction. The goal of these applications is to realize an inline packet processing pipeline to receive packets in GPU memory (without staging copies through CPU memory), process them in parallel with one or more CUDA kernels, and then run inference, evaluate, or send the result of the calculation over the network.

gpu-processing-flow.png

The type of data processing heavily depends on the use case. The goal of this application is to provide a basic layout to reuse in the most common use cases of being able to receive, differentiate and manage the following types of network traffic in multiple queues: UDP, TCP and ICMP.

This application is an enhancement of the use cases presented in the NVIDIA Blog post about DOCA GPUNetIO.

This is a receive-and-process DOCA application, so a packet generator sending packets is required to test the application.

system-design-diagram.png

To launch the application, the PCIe address of the GPU and NIC are needed. Before running the application, ensure that configurations explained in NVIDIA DOCA GPUNetIO Programming Guide and NVIDIA DOCA Flow Programming Guide are respected. For example, make sure to configure huge pages of 1GB (DOCA Flow and DPDK requirement):

Copy
Copied!
            

$ sudo vim /etc/default/grub # Add hugepages size and number along with other options you have in the CMDLINE # GRUB_CMDLINE_LINUX_DEFAULT="default_hugepagesz=1G hugepagesz=1G hugepages=4" $ sudo update-grub $ sudo reboot # After rebooting, check huge pages info $ grep -i huge /proc/meminfo AnonHugePages: 0 kB ShmemHugePages: 0 kB HugePages_Total: 16 HugePages_Free: 15 HugePages_Rsvd: 0 HugePages_Surp: 0 Hugepagesize: 1048576 kB Hugetlb: 16777216 kB


List of items to check before running the application:

  • CUDA Toolkit is installed at 12.1 version or newer
  • doca-gpu and doca-gpu_dev packages are installed
  • gdrcopy installed in /opt/mellanox/gdrcopy:
    • export LD_LIBRARY_PATH=${LD_LIBRARY_PATH}:/opt/mellanox/gdrcopy/src
  • gdrdrv kernel module is active and running on the system
  • nvidia-peermem kernel module is active and running on the system
  • DOCA and DPDK library path are preset in the LD_LIBRARY_PATH environment variable:
    • export LD_LIBRARY_PATH=${LD_LIBRARY_PATH}:/opt/mellanox/dpdk/lib/x86_64-linux-gnu:/opt/mellanox/doca/lib/x86_64-linux-gnu
  • Network card is correctly configured and has a direct PCIe connection with the GPU

If the application is running in a multi-GPU environment, either choose the GPU to use by setting the CUDA_VISIBLE_DEVICES environment variable, or add this simple piece of code in the gpu_packet_processing.c file in the main function right after the doca_argp_start function:

Copy
Copied!
            

int cuda_id; cudaDeviceGetByPCIBusId(&cuda_id, app_cfg.gpu_pcie_addr); cudaFree(0); cudaSetDevice(cuda_id);

The application manages different types of traffic differently, dedicating up to 4 receive queues to each one using DOCA Flow with RSS mode to assign each packet to the right queue. The more queues the application uses, the higher is the degree of parallelism in how receive data is processed and how long it takes.

Tip:

It is highly recommended to use more than one receive queue for 100Gb/s or higher network traffic throughput.

3.1. ICMP Network Traffic

If the network interface used for the application has an IP address, it is possible to ping that interface. ICMP packets are received by a dedicated CUDA kernel (file gpu_kernels/receive_icmp.cu) which:

  1. Receives packets using the DOCA GPUNetIO CUDA warp-level function doca_gpu_dev_eth_rxq_receive_warp.
  2. Checks if the packet is an ICMP echo request.
  3. Forwards the same packet, modifying some header info (e.g., swapping MAC and IP addresses, changing ICMP packet type).
  4. Pushes the modified packet into the send queue using the DOCA GPUNetIO thread-level function doca_gpu_dev_eth_txq_send_enqueue_strong.
  5. Sends the packet using the DOCA GPUNetIO thread-level functions doca_gpu_dev_eth_txq_commit_strong and doca_gpu_dev_eth_txq_push.

Note:

This is not a compute intensive use case so a single CUDA warp with only one receive queue and one send queue is enough to keep up with a decent latency.

By default, the OS CPU ping TTL is set to 64. Therefore, to be sure the GPU is actually replying to ICMP ping requests, TTL is set to 128 in this application.


icmp-network-traffic-diagram.png

Motivations for this use case:

  • Provide an easy tool to check connectivity between packet the generator machine and the DOCA application machine
  • Have a sense of network latency between the two machines using a well-known tool like ping
  • Show an easy way to receive and forward modified packets
  • Provide a warp-level implementation of a CUDA kernel receiving and forwarding traffic

Assuming the network interface to ping has the IP address 192.168.1.1, this is the expected output:

Copy
Copied!
            

$ ping 192.168.1.1 PING 192.168.1.1 (192.168.1.1) 56(84) bytes of data. 64 bytes from 192.168.1.1: icmp_seq=1 ttl=64 time=0.324 ms 64 bytes from 192.168.1.1: icmp_seq=2 ttl=64 time=0.332 ms 64 bytes from 192.168.1.1: icmp_seq=3 ttl=64 time=0.299 ms 64 bytes from 192.168.1.1: icmp_seq=4 ttl=64 time=0.309 ms 64 bytes from 192.168.1.1: icmp_seq=5 ttl=64 time=0.323 ms 64 bytes from 192.168.1.1: icmp_seq=6 ttl=64 time=0.300 ms 64 bytes from 192.168.1.1: icmp_seq=7 ttl=64 time=0.274 ms 64 bytes from 192.168.1.1: icmp_seq=8 ttl=64 time=0.314 ms 64 bytes from 192.168.1.1: icmp_seq=9 ttl=64 time=0.327 ms 64 bytes from 192.168.1.1: icmp_seq=10 ttl=64 time=0.384 ms # At this point, the DOCA application has been started on the 192.168.1.1 interface # TTL becomes 128 as it's the GPU replying to ICMP requests now instead of the OS 64 bytes from 192.168.1.1: icmp_seq=11 ttl=128 time=0.346 ms 64 bytes from 192.168.1.1: icmp_seq=12 ttl=128 time=0.274 ms 64 bytes from 192.168.1.1: icmp_seq=13 ttl=128 time=0.294 ms 64 bytes from 192.168.1.1: icmp_seq=14 ttl=128 time=0.240 ms 64 bytes from 192.168.1.1: icmp_seq=15 ttl=128 time=0.273 ms 64 bytes from 192.168.1.1: icmp_seq=16 ttl=128 time=0.238 ms 64 bytes from 192.168.1.1: icmp_seq=17 ttl=128 time=0.252 ms 64 bytes from 192.168.1.1: icmp_seq=18 ttl=128 time=0.232 ms 64 bytes from 192.168.1.1: icmp_seq=19 ttl=128 time=0.278 ms 64 bytes from 192.168.1.1: icmp_seq=20 ttl=128 time=0.276 ms 64 bytes from 192.168.1.1: icmp_seq=21 ttl=128 time=0.294 ms ......

3.2. UDP Network Traffic

This is the most generic use case of receive-and-analyze packet headers. Designed to keep-up with 100Gb/s of incoming network traffic, the CUDA kernel responsible for the UDP traffic dedicates one CUDA block of 512 CUDA threads (file gpu_kernels/receive_udp.cu) to a different Ethernet UDP receive queue. The data path loop is:

  1. Receive packets using the DOCA GPUNetIO CUDA block-level function doca_gpu_dev_eth_rxq_receive_block.
  2. Each CUDA thread works on a subset of received packets.
  3. DOCA buffer containing the packet is retrieved.
  4. Packet payload is analyzed to differentiate between DNS packets from other UDP generic packets.
  5. Packet payload is wiped-out to ensure that old stale packets are not analyzed again.
  6. Each CUDA block reports to the CPU thread statistics about types of received packets through a DOCA GPUNetIO semaphore.
  7. CPU thread polls on semaphores to retrieve and print the statistics to the console.

udp-network-traffic-diagram.png

The motivation for this use case is mostly to provide an application template to:

  • Receive and analyze packet headers to differentiate across different UDP protocols
  • Report statistics to the CPU through the DOCA GPUNetIO semaphore

Several well-known packet generators can be used to test this mode like T-Rex or DPDK testpmd.

3.3. TCP Network Traffic and HTTP Echo Server

By default, the TCP flow management is the same as UDP: Receive TCP packets and analyze their headers to report to the CPU statistics about the types of received packets.

This is good for passive traffic analyzers or sniffers but sometimes a packet processing application needs to receive packets directly from TCP peers which implies the establishment of a TCP-reliable connection through the 3-way handshake method. Therefore, it is possible to enable TCP “server” mode through the -s command-line flag which enables an “HTTP echo server” mode where the CPU and GPU cooperate to establish a TCP connection and process TCP data packets. Specifically, in this case there are two different sets of receive queues:

  • CPU DPDK receive queues which receive TCP “control” packets (e.g. SYN, FIN or RST)
  • DOCA GPUNetIO receive queues to receive TCP “data” packets

This distinction is possible thanks to DOCA Flow capabilities.

The application’s flow requires CPU and GPU collaboration as described in the following subsections.

3.3.1. Step 1: TCP Connection Establishment

A CPU thread through DPDK queues receives a TCP SYN packet from a remote TCP peer. The CPU thread establishes a TCP reliable connection (replies with a TCP SYN-ACK packet) with the peer and uses DOCA Flow to create a new steering rule to redirect TCP data packets to one of the DOCA GPUNetIO receive queues. The new steering rule excludes control packets (e.g., SYN, FIN or RST).

3.3.2. Step 2: TCP Data Processing

The CUDA kernel responsible for TCP processing receives TCP data packets and performs TCP packet header analysis. If it receives an HTTP GET request, it stores the relevant packet’s info in the next item of a DOCA GPUNetIO semaphore, setting it to READY.

3.3.3. Step 3: HTTP Echo Server

A second CUDA kernel responsible for HTTP processing polls the DOCA GPUNetIO semaphore. Once it detects the update of the next item to READY, it reads the HTTP GET packet info and crafts an HTTP response packet with an HTML page.

If the request is about index.html or contacts.html, the CUDA kernel replies with the appropriate HTML page using a 200 OK code. For all other requests, the it returns a “Page not found” and 404 Error code. HTTP response packets are sent by this second HTTP CUDA kernel using DOCA GPUNetIO.

Note:

Care must be taken to maintain TCP sequence/ack numbers in the packet headers.

3.3.4. Step 4: TCP Connection Closure

If the CPU receives a TCP FIN packet through the DPDK queues, it closes the connection with the remote TCP peer and removes the DOCA Flow rule from the DOCA GPUNetIO queues so the CUDA kernel cannot receive anymore packets from that TCP peer.

tcp-connection-closure-diagram.png

Motivations for this use case:

  • Receiving and analyzing packet headers to differentiate across different TCP protocols
  • Processing TCP packets on GPU in passive mode (sniffing) and active mode (reliable connection)
  • Having a DOCA-DPDK application able to establish a TCP reliable connection without using any OS socket and bypassing kernel routines
  • Having CUDA-kernel-to-CUDA-kernel communication through a DOCA GPUNetIO semaphore
  • Showing how to create and send a packet from scratch with DOCA GPUNetIO

Assuming the network interface used to run the application has the IP address 192.168.1.1 , it is possible to test this HTTP echo server mode using simple tools like curl or wget. Example with curl:

Copy
Copied!
            

$ curl http://192.168.1.1/index.html -ivvv * Trying 192.168.1.1:80... * Connected to 192.168.1.1 (192.168.1.1) port 80 (#0) > GET /index.html HTTP/1.1 > Host: 192.168.1.1 > User-Agent: curl/7.81.0 > Accept: */* > * Mark bundle as not supporting multiuse < HTTP/1.1 200 OK HTTP/1.1 200 OK < Date: Sun, 30 Apr 2023 20:30:40 GMT Date: Sun, 30 Apr 2023 20:30:40 GMT < Content-Type: text/html; charset=UTF-8 Content-Type: text/html; charset=UTF-8 < Content-Length: 158 Content-Length: 158 < Last-Modified: Sun, 30 Apr 2023 22:38:34 GMT Last-Modified: Sun, 30 Apr 2023 22:38:34 GMT < Server: GPUNetIO Server: GPUNetIO < Accept-Ranges: bytes Accept-Ranges: bytes < Connection: keep-alive Connection: keep-alive < Keep-Alive: timeout=5 Keep-Alive: timeout=5 < <html> <head> <title>GPUNetIO index page</title> </head> <body> <p>Hello World, the GPUNetIO server Index page!</p> </body> </html> * Connection #0 to host 192.168.1.1 left intact

This application leverages the following DOCA libraries:

The following explains the application’s flow, highlighting main code blocks and functions:

  1. Parse application argument.
    Copy
    Copied!
                

    doca_argp_init(); register_application_params(); doca_argp_start();

  2. Initialize network device as DOCA device, initialize DPDK, and get device DPDK port ID.
    Copy
    Copied!
                

    init_doca_device();

    Calls rte_eal_init() with empty flags to initialize EAL resources.

  3. Initialize a GPU device, creating a DOCA GPUNetIO handle for it.
    Copy
    Copied!
                

    doca_gpu_create();

  4. Initialize DOCA Flow, starting the DPDK port.
    Copy
    Copied!
                

    init_doca_flow();

    Flags to initialize DOCA Flow are VNF, HW steering, and isolated mode (to prevent the default RSS flows from interfering with the GPUNetIO queues).

  5. Create RX and TX queue related objects (i.e., Ethernet handlers, GPUNetIO handlers, flow rules, semaphores) to manage UDP, TCP and ICMP flows.
    Copy
    Copied!
                

    create_udp_queues(); create_tcp_queues(); create_icmp_queues(); /* Depending on TCP mode (HTTP server or not) properly connect different DOCA Flow pipes */ create_root_pipe();

  6. Allocate generic exit flag. All CUDA kernels periodically poll on this flag. If the CPU set it to 1, CUDA kernels exit from their main loop and return.
    Copy
    Copied!
                

    doca_gpu_mem_alloc(gpu_dev, sizeof(uint32_t), alignment, DOCA_GPU_MEM_GPU_CPU, (void **)&gpu_exit_condition, (void **)&cpu_exit_condition);

  7. Launch CUDA kernels, each on a different stream.
    Copy
    Copied!
                

    kernel_receive_udp(rx_udp_stream, gpu_exit_condition, &udp_queues); kernel_receive_tcp(rx_tcp_stream, gpu_exit_condition, &tcp_queues, app_cfg.http_server); kernel_receive_icmp(rx_icmp_stream, gpu_exit_condition, &icmp_queues); if (app_cfg.http_server) kernel_http_server(tx_http_server, gpu_exit_condition, &tcp_queues, &http_queues);

  8. Launch the CPU thread responsible to poll on DOCA GPUNetIO semaphores and print UDP and TCP stats on the console.
    Copy
    Copied!
                

    rte_eal_remote_launch((void *)stats_core, NULL, current_lcore);

  9. Launch CPU thread responsible for managing TCP 3-way handshake connections.
    Copy
    Copied!
                

    if (app_cfg.http_server) { ... rte_eal_remote_launch(tcp_cpu_rss_func, &tcp_queues, current_lcore); }

  10. Wait for the user to send a signal to quit the application. When this happens, the signal handler function sets the force_quit flag to true which causes the main thread to move forward and set the exit condition to 1.
    Copy
    Copied!
                

    while (DOCA_GPUNETIO_VOLATILE(force_quit) == false); DOCA_GPUNETIO_VOLATILE(*cpu_exit_condition) = 1;

  11. Wait for CUDA kernels to exit and finalize all DOCA Flow and GPUNetIO resources.
    Copy
    Copied!
                

    cudaStreamSynchronize(rx_udp_stream); cudaStreamSynchronize(rx_tcp_stream); cudaStreamSynchronize(rx_icmp_stream); if (app_cfg.http_server) cudaStreamSynchronize(tx_http_server); destroy_flow_queue(); doca_gpu_destroy();

  1. Refer to the following documents:
  2. The gpu_packet_processing binary is located under /opt/mellanox/doca/applications/gpu_packet_processing/bin/doca_gpu_packet_processing. To build all the applications together, run:
    Copy
    Copied!
                

    cd /opt/mellanox/doca/applications/ meson build ninja -C build

  3. To build only this application:
    1. Edit the following flags in /opt/mellanox/doca/applications/meson_options.txt:
      • Set enable_all_applications to false
      • Set enable_gpu_support to true
      • Set enable_gpu_packet_processing to true
    2. Run the commands in step 2.
      Note:

      doca_gpu_packet_processing is created under ./build/gpu_packet_processing/src/.

    Application usage:

    Copy
    Copied!
                

    Usage: doca_gpu_packet_processing [DOCA Flags] [Program Flags] DOCA Flags: -h, --help Print a help synopsis -v, --version Print program version information -l, --log-level Set the log level for the program <CRITICAL=20, ERROR=30, WARNING=40, INFO=50, DEBUG=60> Program Flags: -g, --gpu <GPU PCIe address> GPU PCIe address to be used by the app -n, --nic <NIC PCIe address> DOCA device PCIe address used by the app -q, --queue <GPU receive queues> DOCA GPUNetIO receive queue per flow -s, --httpserver <Enable GPU HTTP server> Enable GPU HTTP server mode


  4. To run the application on the host, assuming a GPU PCIe address ca:00.0 and NIC PCIe address 17:00.0 with 2 GPUNetIO receive queues:
    Copy
    Copied!
                

    doca_gpu_packet_processing -n 17:00.0 -g ca:00.0 -q 2

    Note:

    Refer to section “Running DOCA Application on Host” in the NVIDIA DOCA Virtual Functions User Guide.

Refer to NVIDIA DOCA Arg Parser Programming Guide for more information.

Flag Type Short Flag Long Flag/JSON Key Description
DPDK flags a devices Adds a PCIe device into the list of devices to probe.
l core-list List of cores to run on.
General flags l log-level Sets the log level for the application:
  • CRITICAL=20
  • ERROR=30
  • WARNING=40
  • INFO=50
  • DEBUG=60
v version Prints program version information.
h help Prints a help synopsis.
Program flags g gpu GPU PCIe address in <bus>:<device>.<function> format. This can be obtained using the nvidia-smi or lspci commands.
n nic Network card port PCIe address in <bus>:<device>.<function> format. This can be obtained using the lspci command.
q queue Number of receive queues to use in the example. Default is 1, maximum allowed is 4.
s httpserver Enable the TCP HTTP server mode. With this flag, TCP packets are not received by GPUNetIO as regular sniffer as it requires a TCP 3-way handshake to establish a reliable connection first.


Notice

This document is provided for information purposes only and shall not be regarded as a warranty of a certain functionality, condition, or quality of a product. NVIDIA Corporation nor any of its direct or indirect subsidiaries and affiliates (collectively: “NVIDIA”) make no representations or warranties, expressed or implied, as to the accuracy or completeness of the information contained in this document and assume no responsibility for any errors contained herein. NVIDIA shall have no liability for the consequences or use of such information or for any infringement of patents or other rights of third parties that may result from its use. This document is not a commitment to develop, release, or deliver any Material (defined below), code, or functionality.

NVIDIA reserves the right to make corrections, modifications, enhancements, improvements, and any other changes to this document, at any time without notice.

Customer should obtain the latest relevant information before placing orders and should verify that such information is current and complete.

NVIDIA products are sold subject to the NVIDIA standard terms and conditions of sale supplied at the time of order acknowledgement, unless otherwise agreed in an individual sales agreement signed by authorized representatives of NVIDIA and customer (“Terms of Sale”). NVIDIA hereby expressly objects to applying any customer general terms and conditions with regards to the purchase of the NVIDIA product referenced in this document. No contractual obligations are formed either directly or indirectly by this document.

NVIDIA products are not designed, authorized, or warranted to be suitable for use in medical, military, aircraft, space, or life support equipment, nor in applications where failure or malfunction of the NVIDIA product can reasonably be expected to result in personal injury, death, or property or environmental damage. NVIDIA accepts no liability for inclusion and/or use of NVIDIA products in such equipment or applications and therefore such inclusion and/or use is at customer’s own risk.

NVIDIA makes no representation or warranty that products based on this document will be suitable for any specified use. Testing of all parameters of each product is not necessarily performed by NVIDIA. It is customer’s sole responsibility to evaluate and determine the applicability of any information contained in this document, ensure the product is suitable and fit for the application planned by customer, and perform the necessary testing for the application in order to avoid a default of the application or the product. Weaknesses in customer’s product designs may affect the quality and reliability of the NVIDIA product and may result in additional or different conditions and/or requirements beyond those contained in this document. NVIDIA accepts no liability related to any default, damage, costs, or problem which may be based on or attributable to: (i) the use of the NVIDIA product in any manner that is contrary to this document or (ii) customer product designs.

No license, either expressed or implied, is granted under any NVIDIA patent right, copyright, or other NVIDIA intellectual property right under this document. Information published by NVIDIA regarding third-party products or services does not constitute a license from NVIDIA to use such products or services or a warranty or endorsement thereof. Use of such information may require a license from a third party under the patents or other intellectual property rights of the third party, or a license from NVIDIA under the patents or other intellectual property rights of NVIDIA.

Reproduction of information in this document is permissible only if approved in advance by NVIDIA in writing, reproduced without alteration and in full compliance with all applicable export laws and regulations, and accompanied by all associated conditions, limitations, and notices.

THIS DOCUMENT AND ALL NVIDIA DESIGN SPECIFICATIONS, REFERENCE BOARDS, FILES, DRAWINGS, DIAGNOSTICS, LISTS, AND OTHER DOCUMENTS (TOGETHER AND SEPARATELY, “MATERIALS”) ARE BEING PROVIDED “AS IS.” NVIDIA MAKES NO WARRANTIES, EXPRESSED, IMPLIED, STATUTORY, OR OTHERWISE WITH RESPECT TO THE MATERIALS, AND EXPRESSLY DISCLAIMS ALL IMPLIED WARRANTIES OF NONINFRINGEMENT, MERCHANTABILITY, AND FITNESS FOR A PARTICULAR PURPOSE. TO THE EXTENT NOT PROHIBITED BY LAW, IN NO EVENT WILL NVIDIA BE LIABLE FOR ANY DAMAGES, INCLUDING WITHOUT LIMITATION ANY DIRECT, INDIRECT, SPECIAL, INCIDENTAL, PUNITIVE, OR CONSEQUENTIAL DAMAGES, HOWEVER CAUSED AND REGARDLESS OF THE THEORY OF LIABILITY, ARISING OUT OF ANY USE OF THIS DOCUMENT, EVEN IF NVIDIA HAS BEEN ADVISED OF THE POSSIBILITY OF SUCH DAMAGES. Notwithstanding any damages that customer might incur for any reason whatsoever, NVIDIA’s aggregate and cumulative liability towards customer for the products described herein shall be limited in accordance with the Terms of Sale for the product.

Trademarks

NVIDIA, the NVIDIA logo, and Mellanox are trademarks and/or registered trademarks of Mellanox Technologies Ltd. and/or NVIDIA Corporation in the U.S. and in other countries. The registered trademark Linux® is used pursuant to a sublicense from the Linux Foundation, the exclusive licensee of Linus Torvalds, owner of the mark on a world¬wide basis. Other company and product names may be trademarks of the respective companies with which they are associated.

Copyright

© 2023 NVIDIA Corporation & affiliates. All rights reserved.

© Copyright 2023, NVIDIA. Last updated on Nov 13, 2023.