How Do I Enable NIC Multi-Queue for an Image?

Scenarios

Network I/O bandwidth can keep increasing to the point where a single vCPU cannot process all of the NIC interrupts. NIC multi-queue allows multiple vCPUs to process NIC interrupts, thereby improving network PPS and I/O performance.

ECSs Supporting NIC Multi-Queue

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

• For details about the ECS flavors that support NIC multi-queue, see section "Instances" in Elastic Cloud Server User Guide.
  

  If there are more than 1 NIC queue, NIC multi-queue is supported.

• Only KVM ECSs support NIC multi-queue.
• The Linux public images listed in Table 1 support NIC multi-queue.
  

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

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

Operation Instructions

Assume that an ECS has the required specifications and virtualization type.

• If the ECS was created from a public image listed in ECSs Supporting NIC Multi-Queue, NIC multi-queue is enabled on the ECS by default. You do not need to enable NIC multi-queue manually.
• If the ECS was created from an external image file with an OS listed in ECSs Supporting NIC Multi-Queue, you may need to perform the following operations to enable NIC multi-queue:
  1. Register the External Image File as a Private Image.
  2. Enable NIC Multi-Queue for the Image.
  3. Create an ECS from the Private Image.
  4. Enable NIC Multi-Queue on the ECS.

Register the External Image File as a Private Image

Register the external image file as a private image. For details, see Registering an External Image File as a Private Image.

Enable NIC Multi-Queue for the Image

Windows has not commercially supported NIC multi-queue. If you enable NIC multi-queue for a Windows image, an ECS created from such an image may take longer than normal to start.

Use any of the following methods to enable NIC multi-queue for an image:

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

1. Obtain a token. For details, see Calling APIs > Authentication in Image Management Service API Reference.
2. Call an API to update image information. For details, 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.

   Set Content-Type to 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
            } 
    ]

Create an ECS from the Private Image

Enable NIC Multi-Queue on the ECS

KVM ECSs running Windows use private images to support NIC multi-queue.

For Linux ECSs, which run CentOS 7.4 as an example, perform the following operations 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

   Example:

   [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 to 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 1 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 fails to meet your expectations, manually configure interrupt affinity on the target ECS.

   The detailed operations are as follows:

   Run the following script so that each ECS vCPU responds the interrupt requests initialized by one queue. That is, 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