Enabling NIC Multi-Queue

Scenarios

With the increase of network I/O bandwidth, single-core CPUs face bottlenecks in handling network interrupts. NIC multi-queue assigns interrupts to different CPUs for higher packets per second (PPS) and bandwidth.

Support of NIC Multi-Queue

NIC multi-queue can be enabled on an ECS only when the ECS specifications, virtualization type, and image OS meet the requirements described in this section.

• For details about the ECS specifications that support NIC multi-queue, see x86 ECS Specifications.
  

  If the number of NIC queues is greater than 1, NIC multi-queue is supported.

• The virtualization type must be KVM.
• The Linux public images listed in Table 2 support NIC multi-queue.
  

  • Windows public images have not supported NIC multi-queue. If you enable NIC multi-queue in a Windows public image, starting an ECS created using such an image may be slow.
  • It is a good practice to upgrade the kernel version of the Linux ECS to 2.6.35 or later. Otherwise, NIC multi-queue is not supported.

    Run the uname -r command to obtain the kernel version. If the kernel version is earlier than 2.6.35, contact technical support to upgrade the kernel.

Importing the External Image File to the IMS Console

For details, see "Registering an Image File as a Private Image" in Image Management Service User Guide.

Setting NIC Multi-Queue for the Image

Windows OSs have not commercially supported NIC multi-queue. If you enable NIC multi-queue in a Windows image, starting an ECS created using such an image may be slow.

Use one of the following methods to set the NIC multi-queue attribute:

Method 3: Add hw_vif_multiqueue_enabled to an image through the API.

1. For instructions about how to obtain the token, see Calling APIs > Authentication in Image Management Service API Reference.
2. For instructions about how to call an API to update image information, see "Updating Image Information (Native OpenStack API)" in Image Management Service API Reference.
3. Add X-Auth-Token to the request header.

   The value of X-Auth-Token is the token obtained in step 1.

4. Add Content-Type to the request header.

   The value of Content-Type is application/openstack-images-v2.1-json-patch.

   The request URI is in the following format:

   PATCH /v2/images/{image_id}

   The request body is as follows:

   [       
            { 
             "op":"add",
             "path":"/hw_vif_multiqueue_enabled", 
             "value": "true" 
            } 
    ]

   Figure 1 shows an example request body for modifying the NIC multi-queue attribute.

   Figure 1 Example request body
   

Creating an ECS Using a Private Image

Enabling NIC Multi-Queue

KVM Windows ECSs use private images to support NIC multi-queue. For details, see "How Do I Set NIC Multi-queue Feature of an Image?" in Image Management Service User Guide.

This section uses a Linux ECS running CentOS 7.4 as an example to describe how to enable NIC multi-queue.

1. Enable NIC multi-queue.
   1. Log in to the ECS.
   2. Run the following command to obtain the number of queues supported by the NIC and the number of queues with NIC multi-queue enabled:

      ethtool -l NIC

   3. Run the following command to configure the number of queues used by the NIC:

      ethtool -L NIC combined Number of queues

   An example is provided as follows:

   [root@localhost ~]# ethtool -l eth0  #View the number of queues used by NIC eth0.
   Channel parameters for eth0:
   Pre-set maximums:
   RX:               0
   TX:               0
   Other:                  0
   Combined: 4  #Indicates that a maximum of four queues can be enabled for the NIC.
   Current hardware settings:
   RX:               0
   TX:               0
   Other:                  0
   Combined: 1 #Indicates that one queue has been enabled.
   
   [root@localhost ~]# ethtool -L eth0 combined 4 #Enable four queues on NIC eth0.

2. (Optional) Enable irqbalance so that the system automatically allocates NIC interrupts on multiple vCPUs.
   1. Run the following command to enable irqbalance:

      service irqbalance start

   2. Run the following command to view the irqbalance status:

      service irqbalance status

      If the Active value in the command output contains active (running), irqbalance has been enabled.

      Figure 2 Enabled irqbalance
      

3. (Optional) Enable interrupt binding.

   Enabling irqbalance allows the system to automatically allocate NIC interrupts, improving network performance. If the improved network performance still fails to meet your requirements, manually configure interrupt affinity on the ECS.

   To do so, perform the following operations:

   Configure the following script so that one ECS vCPU serves the interrupt requests initialized by one queue. One queue corresponds to one interrupt, and one interrupt binds to one vCPU.

   #!/bin/bash
   service irqbalance stop
   
   eth_dirs=$(ls -d /sys/class/net/eth*)
   if [ $? -ne 0 ];then
       echo "Failed to find eth*  , sleep 30" >> $ecs_network_log
       sleep 30
       eth_dirs=$(ls -d /sys/class/net/eth*)
   fi
   
   for eth in $eth_dirs
   do
       cur_eth=$(basename $eth)
       cpu_count=`cat /proc/cpuinfo| grep "processor"| wc -l`
       virtio_name=$(ls -l /sys/class/net/"$cur_eth"/device/driver/ | grep pci |awk {'print $9'})
   
       affinity_cpu=0
       virtio_input="$virtio_name""-input"
       irqs_in=$(grep "$virtio_input" /proc/interrupts | awk -F ":" '{print $1}')
       for irq in ${irqs_in[*]}
       do
           echo $((affinity_cpu%cpu_count)) > /proc/irq/"$irq"/smp_affinity_list
           affinity_cpu=$[affinity_cpu+2]
       done
   
       affinity_cpu=1
       virtio_output="$virtio_name""-output"
       irqs_out=$(grep "$virtio_output" /proc/interrupts | awk -F ":" '{print $1}')
       for irq in ${irqs_out[*]}
       do
           echo $((affinity_cpu%cpu_count)) > /proc/irq/"$irq"/smp_affinity_list
           affinity_cpu=$[affinity_cpu+2]
       done
   done

4. (Optional) Enable XPS and RPS.

   XPS allows the system with NIC multi-queue enabled to select a queue by vCPU when sending a data packet.

   #!/bin/bash
   # enable XPS feature
   cpu_count=$(grep -c processor /proc/cpuinfo)
   dec2hex(){
     echo $(printf "%x" $1)
   }
   eth_dirs=$(ls -d /sys/class/net/eth*)
   if [ $? -ne 0 ];then
       echo "Failed to find eth* , sleep 30" >> $ecs_network_log
       sleep 30
       eth_dirs=$(ls -d /sys/class/net/eth*)
   fi
   for eth in $eth_dirs
   do
       cpu_id=1
       cur_eth=$(basename $eth)
       cur_q_num=$(ethtool -l $cur_eth | grep -iA5 current | grep -i combined | awk {'print $2'})
       for((i=0;i<cur_q_num;i++))
       do
           if [ $i -eq $cpu_count ];then
               cpu_id=1
           fi
           xps_file="/sys/class/net/${cur_eth}/queues/tx-$i/xps_cpus"
           rps_file="/sys/class/net/${cur_eth}/queues/rx-$i/rps_cpus"
           cpuset=$(dec2hex "$cpu_id")
           echo $cpuset > $xps_file
           echo $cpuset > $rps_file
           let cpu_id=cpu_id*2
       done
   done