Overview

This wiki page describes how to tune Mellanox OpenStack on Ubuntu 12.04 servers

Prerequisites

A running OpenStack environment installed with the ML2 plugin on top of OVS.
All nodes equipped with Mellanox ConnectX®-3 Network Adapter (http://www.mellanox.com/page/products_dyn?product_family=119)
Mellanox OFED 2.2 or greater installed on all nodes. Please refer to Mellanox website for the latest OFED: http://www.mellanox.com/page/products_dyn?product_family=26&mtag=linux_sw_drivers
SR-IOV enabled on all compute nodes. For more information, please refer to Mellanox Community click here.
The software package iproute2 - (http://www.linuxfoundation.org/collaborate/workgroups/networking/iproute2 ) installed on all Compute nodes
VLANs configured on the ports in the switch.

InfiniBand Network

OpenSM Node

To configure OpenSM node:

1. Edit the file /etc/opensm/partitions.conf by adding the following line:

management=0xffff,ipoib, sl=0, defmember=both : ALL, ALL_SWITCHES=full,SELF=full;

Create the file if it does not exist.

For each VLAN value in the VLAN range used for VLAN network segmentation, add the following line in the file /etc/opensm/partitions.conf (for vlan-pkey association):

vlan<decimal-value>=0x<hexadecimal-value>, ipoib, sl=0, defmember=full: ALL;

In this manual, the VLAN range is 2-10 (mlnx_agent.conf /ml2_conf.ini). Therefore, add the following lines:

vlan1=0x1, ipoib, sl=0, defmember=full : ALL;
vlan2=0x2, ipoib, sl=0, defmember=full : ALL;
vlan3=0x3, ipoib, sl=0, defmember=full : ALL;
vlan4=0x4, ipoib, sl=0, defmember=full : ALL;
vlan5=0x5, ipoib, sl=0, defmember=full : ALL;
vlan6=0x6, ipoib, sl=0, defmember=full : ALL;
vlan7=0x7, ipoib, sl=0, defmember=full : ALL;
vlan8=0x8, ipoib, sl=0, defmember=full : ALL;
vlan9=0x9, ipoib, sl=0, defmember=full : ALL;
vlan10=0xa, ipoib, sl=0, defmember=full : ALL;

2. Modify the following line in the file /etc/opensm/opensm.conf from FALSE to TRUE:

allow_both_pkeys TRUE

3. Restart OpenSM:

service opensmd restart

4. Verify that all ports on all nodes have their ib interface (i.e. ib0) active, run ibdev2netdev on all hosts, and make sure that the port is up:

mlx4_0 port 1 ==> ib0 (Up)

If the port is down, try bringing it up by running “ifconfig ib0 up”, and make sure your switch ports are of type “InfiniBand”, and that all nodes can reach the OpenSM node by running “sminfo”.

Controller Node

To configure the Controller node:

1. Ensure that the file /etc/nova/nova.conf includes the following line:

security_group_api=nova

2. If the security group value is changed, restart the Nova services:

# for i in $(initctl list | grep '^nova' | awk '{print $1}'); do service $i restart; done

Neutron Server Node

NOTE: The Neutron Server can sometimes be part of the Controller node.

To configure the Neutron Server node:

1. Edit /etc/neutron/neutron.conf and make sure Neutron is using ML2 plugin:

[DEFAULT]
core_plugin = neutron.plugins.ml2.plugin.Ml2Plugin

2. Edit the ML2 configuration file /etc/neutron/plugins/ml2/ml2_conf.ini as follows:

[ml2]
type_drivers = vlan,flat
tenant_network_types = vlan 
mechanism_drivers = linuxbridge,mlnx
 
[ml2_type_vlan]
network_vlan_ranges=<physical_network>:<min_vlan>:<max_vlan> //(i.e. default:2:10)

[securitygroup]
enable_security_group = True

[eswitch]
vnic_type = hostdev
apply_profile_patch = True

3. Install the required Deb for the Mellanox Neutron plugin (from Ubuntu’s repository):

apt-get install neutron-plugin-mlnx-agent

4. Edit /etc/neutron/plugins/mlnx/mlnx_conf.ini to match your environment:

[mlnx]
network_vlan_ranges = <physical_network>:<min_vlan:<max_vlan> //(i.e. default:2:10)
tenant_network_type = vlan

[eswitch]
physical_interface_mappings = <physical_network>:<ib interface>  //(i.e. default:ib0)
vnic_type = hostdev

[agent]
support_old_agents = True

5. Restart Neutron Server:

/etc/init.d/neutron-server start

Compute Nodes

Installation

1. If you didn't download Mellanox OpenStack debian files download it :

http://www.mellanox.com/downloads/solutions/openstack/

2. Install the eswitchd debian files:

  dpkg -i eswitchd_0.8-1_amd64.deb
  dpkg -i python-mlnxvif_0.6-1_all.deb
  apt-get install -f # fix dependencies
  apt-get install -y neutron-plugin-mlnx-agent

3. In case you would like to use Ethernet in para-virtualized mode the VIF driver is already included in Nova package. Otherwise, Install Mellanox VIF driver (Make sure nova is installed on your server)

Configuration

1. Configure /etc/eswitchd/eswitchd.conf if needed

Please Refer to Mellanox Community for the eSwitchd installation notes ( click here)

2. Modify /etc/nova/nova.conf

  compute_driver=nova.virt.libvirt.driver.LibvirtDriver
  libvirt_vif_driver=mlnxvif.vif.MlxEthVIFDriver
  security_group_api=nova
  connection_type=libvirt

In case you didn't install the Mellanox VIF driver, and you plan to use Ethernet only in paravirtualized ,mode: change the following:

  libvirt_vif_driver=nova.virt.libvirt.vif.LibvirtGenericVIFDriver

3. Modify the /etc/neutron/plugins/mlnx/mlnx_conf.ini file to reflect your environment.

 [AGENT]
 rpc_support_old_agents = True
 [ESWITCH]
  physical_interface_mapping -  the network_interface_mappings maps each physical network name to the physical interface (on top of Mellanox Adapter) connecting the node to that physical network. The format of this paramter is:     <fabric name>:< PF name> (Only releavant on Compute node). PF Name can either be the PF (Physical Function) Name or 'autoeth' for automatic Ethernet configuration,'autoib' for automatic Infiniband configuration.The default is  "default:autoeth".  
  vnic_type - type of VM network interface: 'mlnx_direct' or 'hostdev'  according to libvirt terminology.
       hostdev: this is the traditional method of assigning any generic PCI device to a guest (SR-IOV).
       mlnx_direct: this is a method to provide macvtap device on top of the PCI device (default).
       bridge:  - When using Linux Bridge Plugin on top of eIPoIB device
  daemon_endpoint - 'tcp://127.0.0.1:60001'

For a plugin configuration file example (Havana), please refer to Mellanox config *ini file (click here)

4. (Optional) Modify /etc/init/neutron-plugin-mlnx-agent.conf as described in LP bug #1257732

5. Check that ebrctl is listed somewhere in /etc/nova/rootwrap.d/*, otherwise add this to /etc/nova/rootwrap.d/compute.filters file in the [Filters] section.

 ebrctl: CommandFilter, ebrctl, root

Start Services

1. Restart Nova.

 #/etc/init.d/nova-compute restart

2. Start eswitch Daemon

  #/etc/init.d/eswitchd start

3. Start the Neutron agent

 #/etc/init.d/neutron-plugin-mlnx-agent start

Note: eswitch Daemon should be running before the Neutron Agent is started.

Network Node

1. Install the Neutron Linux bridge plugin, the Neutron DHCP agent and the Neutron L3 agent:

 # apt-get install neutron-plugin-linuxbridge-agent neutron-dhcp-agent neutron-l3-agent

2. Copy /etc/neutron/neutron.conf from the Compute node.

3. Change the following configuration of the ini file (/etc/neutron/plugins/linuxbridge/linuxbridge_conf.ini). The "default" in the example is the name of the physical network as configured in /etc/neutron/plugins/mlnx/mlnx_conf.ini .

  [linux_bridge]  
  physical_interface_mappings = default:eth2

4. Configure the DHCP agent according to the following guidelines

Update the following file: /etc/neutron/dhcp_agent.ini with:

  [DEFAULT] 
 interface_driver = neutron.agent.linux.interface.BridgeInterfaceDriver

For additional information refer to the OpenStack documentation

5. Configure the L3 agent according to the following guidelines

Update the following file: /etc/neutron/l3_agent.ini with:

  [DEFAULT] 
 interface_driver = neutron.agent.linux.interface.BridgeInterfaceDriver