5. Virtio_user for Container Networking

Container becomes more and more popular for strengths, like low overhead, fast boot-up time, and easy to deploy, etc. How to use DPDK to accelerate container networking becomes a common question for users. There are two use models of running DPDK inside containers, as shown in Fig. 5.4.

Fig. 5.4 Use models of running DPDK inside container

This page will only cover aggregation model.

5.1. Overview

The virtual device, virtio-user, with unmodified vhost-user backend, is designed for high performance user space container networking or inter-process communication (IPC).

The overview of accelerating container networking by virtio-user is shown in Fig. 5.5.

Fig. 5.5 Overview of accelerating container networking by virtio-user

Different virtio PCI devices we usually use as a para-virtualization I/O in the context of QEMU/VM, the basic idea here is to present a kind of virtual devices, which can be attached and initialized by DPDK. The device emulation layer by QEMU in VM’s context is saved by just registering a new kind of virtual device in DPDK’s ether layer. And to minimize the change, we reuse already-existing virtio PMD code (driver/net/virtio/).

Virtio, in essence, is a shm-based solution to transmit/receive packets. How is memory shared? In VM’s case, qemu always shares the whole physical layout of VM to vhost backend. But it’s not feasible for a container, as a process, to share all virtual memory regions to backend. So only those virtual memory regions (aka, hugepages initialized in DPDK) are sent to backend. It restricts that only addresses in these areas can be used to transmit or receive packets.

5.2. Sample Usage

Here we use Docker as container engine. It also applies to LXC, Rocket with some minor changes.

  1. Compile DPDK.

    make install RTE_SDK=`pwd` T=x86_64-native-linuxapp-gcc
    
  2. Write a Dockerfile like below.

    cat <<EOT >> Dockerfile
    FROM ubuntu:latest
    WORKDIR /usr/src/dpdk
    COPY . /usr/src/dpdk
    ENV PATH "$PATH:/usr/src/dpdk/x86_64-native-linuxapp-gcc/app/"
    EOT
    
  3. Build a Docker image.

    docker build -t dpdk-app-testpmd .
    
  4. Start a testpmd on the host with a vhost-user port.

    $(testpmd) -c 0x3 -n 4 --socket-mem 1024,1024 \
        --vdev 'eth_vhost0,iface=/tmp/sock0' --no-pci -- -i
    
  5. Start a container instance with a virtio-user port.

    docker run -i -t -v /tmp/sock0:/var/run/usvhost \
        -v /dev/hugepages:/dev/hugepages \
        dpdk-app-testpmd testpmd -c 0xc -n 4 -m 1024 --no-pci \
        --vdev=virtio_user0,path=/var/run/usvhost \
        -- -i --txqflags=0xf00 --disable-hw-vlan
    

Note: If we run all above setup on the host, it’s a shm-based IPC.

5.3. Limitations

We have below limitations in this solution:
  • Cannot work with –huge-unlink option. As we need to reopen the hugepage file to share with vhost backend.
  • Cannot work with –no-huge option. Currently, DPDK uses anonymous mapping under this option which cannot be reopened to share with vhost backend.
  • Cannot work when there are more than VHOST_MEMORY_MAX_NREGIONS(8) hugepages. In another word, do not use 2MB hugepage so far.
  • Applications should not use file name like HUGEFILE_FMT (“%smap_%d”). That will bring confusion when sharing hugepage files with backend by name.
  • Root privilege is a must. DPDK resolves physical addresses of hugepages which seems not necessary, and some discussions are going on to remove this restriction.