← Back to team overview

kernel-packages team mailing list archive

[Bug 1201869] Re: poor networking throughput through veth interfaces

 

** Description changed:

  SRU Justification:
  
  Impact:
  Users of the 3.2/3.5/3.8 series kernel will have poor network throughput when using OpenStack Neutron depending on their setup.
  
  Fix:
  These upstream patches are necessary to fix the issue:
  
  2681128f0ced8aa4e66f221197e183cc16d244fe
  8093315a91340bca52549044975d8c7f673b28a1
  d0e2c55e7c940a3ee91e9e23a2683b593690f1e9
  2efd32ee1b60b0b31404ca47c1ce70e5a5d24ebc
  f45a5c267da35174e22cec955093a7513dc1623d
- 
- Minimally 8093315a91340bca52549044975d8c7f673b28a1 provides the proper
- features to increase performance dramatically.
  
  Testcase:
  Setup OpenStack Neutron. Test throughput between internal and external nodes.
  
  The following explains an example vlan+namespace configuration:
  Internal Node: [10.x.x.2]->eth2.123->br123->tap123->qr-123[10.x.x.1] <--- netns: qrouter-123
  netns: qrouter-123 ---> qg-234[10.x.y.1]->tap234->br234->eth2.234->External Node[10.x.y.2]
  Where:
  1) tap123+qr-123 and tap234+qg-234 are veth pairs
  2) qr-123 and qg-234 reside inside the qrouter-123 namespace
  
  Another testcase without Openstack:
  
  * create two vms: (vm1, vm2), install iperf on those machines
  * connect vms via an isolated bridge
  * measure baseline performance
    - iperf -s # on machine 1
    - iperf -t 60 -l 4M -c <machine 1 IP> # on machine 2
  * create veth pairs between vms using attached script:
    - ./setup-lp1201869.sh vm1
    - ./setup-lp1201869.sh vm2
  * In the VM's setup static IPs
    - sudo ifconfig eth0 10.10.10.1/24 up #vm1
    - sudo ifconfig eth0 10.10.10.2/24 up #vm2
  * measure performance now
  * we expect this to be close to the existing performance
  
  --
  
  OpenStack Neutron does IP forwarding through a network namespace. A veth
  pair is used to connect into the namespace. The veth pair appears to be
  the bottleneck, independent of network namespace. In newer versions of
  Linux (Ubuntu-3.9.0-7.15 / v3.9-rc1 and greater) throughput is much
  higher by almost 3 times. For example with some testing throughput is
  3.5 Gbps in pre 3.9-rc1 versions and 9.1 Gbps with these patches
  applied.
  
  This has been confirmed on kernels from 3.5.x-3.8.x. (Quantal and Raring
  lts backports)

-- 
You received this bug notification because you are a member of Kernel
Packages, which is subscribed to linux in Ubuntu.
https://bugs.launchpad.net/bugs/1201869

Title:
  poor networking throughput through veth interfaces

Status in “linux” package in Ubuntu:
  Fix Released
Status in “linux” source package in Precise:
  In Progress
Status in “linux” source package in Quantal:
  Fix Released
Status in “linux” source package in Raring:
  Fix Released

Bug description:
  SRU Justification:

  Impact:
  Users of the 3.2/3.5/3.8 series kernel will have poor network throughput when using OpenStack Neutron depending on their setup.

  Fix:
  These upstream patches are necessary to fix the issue:

  2681128f0ced8aa4e66f221197e183cc16d244fe
  8093315a91340bca52549044975d8c7f673b28a1
  d0e2c55e7c940a3ee91e9e23a2683b593690f1e9
  2efd32ee1b60b0b31404ca47c1ce70e5a5d24ebc
  f45a5c267da35174e22cec955093a7513dc1623d

  Testcase:
  Setup OpenStack Neutron. Test throughput between internal and external nodes.

  The following explains an example vlan+namespace configuration:
  Internal Node: [10.x.x.2]->eth2.123->br123->tap123->qr-123[10.x.x.1] <--- netns: qrouter-123
  netns: qrouter-123 ---> qg-234[10.x.y.1]->tap234->br234->eth2.234->External Node[10.x.y.2]
  Where:
  1) tap123+qr-123 and tap234+qg-234 are veth pairs
  2) qr-123 and qg-234 reside inside the qrouter-123 namespace

  Another testcase without Openstack:

  * create two vms: (vm1, vm2), install iperf on those machines
  * connect vms via an isolated bridge
  * measure baseline performance
    - iperf -s # on machine 1
    - iperf -t 60 -l 4M -c <machine 1 IP> # on machine 2
  * create veth pairs between vms using attached script:
    - ./setup-lp1201869.sh vm1
    - ./setup-lp1201869.sh vm2
  * In the VM's setup static IPs
    - sudo ifconfig eth0 10.10.10.1/24 up #vm1
    - sudo ifconfig eth0 10.10.10.2/24 up #vm2
  * measure performance now
  * we expect this to be close to the existing performance

  --

  OpenStack Neutron does IP forwarding through a network namespace. A
  veth pair is used to connect into the namespace. The veth pair appears
  to be the bottleneck, independent of network namespace. In newer
  versions of Linux (Ubuntu-3.9.0-7.15 / v3.9-rc1 and greater)
  throughput is much higher by almost 3 times. For example with some
  testing throughput is 3.5 Gbps in pre 3.9-rc1 versions and 9.1 Gbps
  with these patches applied.

  This has been confirmed on kernels from 3.5.x-3.8.x. (Quantal and
  Raring lts backports)

To manage notifications about this bug go to:
https://bugs.launchpad.net/ubuntu/+source/linux/+bug/1201869/+subscriptions


References