Open Stack

+1 we need more net testing pre deployment to recommend usable network
settings the first time. Settings like mtu gro gso need to be considered
too
On Dec 16, 2014 8:30 AM, "Sergey Vasilenko" <svasilenko at mirantis.com> wrote:

> Guys, it's a big and complicated architecture issue.
>
> Issue, like this was carefully researched about month ago (while P***)
>
> root case of issue:
>
>    - Now we use OVS for build virtual network topology on each node.
>    - OVS has performance degradation while pass huge of small network
>    packets.
>    - We can’t abandon using OVS entirely and forever, because it's a most
>    popular Neutron solution.
>    - We can’t abandon using OVS partial now, because low-level modules
>    don’t ready yet for this. I start blueprint (
>    https://blueprints.launchpad.net/fuel/+spec/l23network-refactror-to-provider-based-resources)
>    for aim possibility of combine using OVS for Neutron purposes and don't use
>    it for management, storage, etc... purposes.
>
> We, together with L2 support team, Neutron team, and another network
> experts make tuning one of existing production-like env after deployment
> and achieve following values on bonds of two 10G cards:
>
>    - vm-to-vm speed (on different compute nodes): 2.56 Gbits/sec (GRE
>    segmentation)
>    - node-to-node speed: 17.6 Gbits/s
>
> This values closely near with theoretical maximum for OVS 1.xx with GRE.
> Some performance improvements may also achieved by upgrading open vSwitch
> to the latest LTS (2.3.1 at this time) branch and using "megaflow" feature (
> http://networkheresy.com/2014/11/13/accelerating-open-vswitch-to-ludicrous-speed/
> ).
>
>
> After this research we concluded:
>
>
>    - OVS can't pass huge of small packages without network performance
>    degradation
>    - for fix this we should re-design network topology on env nodes
>    - even re-designed network topology can't fix this issue at all. Some
>    network parameters, like mtu, disabling offloading for NICs, buffers,
>    etc... can be tuned only on real environment.
>
>
> My opinion — in FUEL we should add new (or extend existing
> network-checker) component.  This component should testing network
> performance on real customer’s pre-configured env by different (already
> defined) performance test cases and recommend better setup BEFORE main
> deployment cycle run.
>
> /sv
>
> _______________________________________________
> OpenStack-dev mailing list
> OpenStack-dev at lists.openstack.org
> http://lists.openstack.org/cgi-bin/mailman/listinfo/openstack-dev
>
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.openstack.org/pipermail/openstack-dev/attachments/20141231/19a7d043/attachment.html>