Neutron’s ovs_lib uses the Open vSwitch CLI command ovs-vsctl to perform basic vSwitch CRUD operations. This is almost an order of magnitude slower than using either direct OVSDB protocol commands, or using the Open vSwitch Python API.
First, an interface for interacting with with the OVSDB should be defined as an abstract base class. Then, an implementation of this interface using ovs-vsctl calls should be made and ovs_lib should make all ovsdb-related calls through this new implementation.
Next, an implementation of the OVSDB interface should be implemented using the Open vSwitch IDL Python library. A configuration option will be added to allow selection of the OVSDB interface to use. It is necessary to allow this choice as there are differences in how privileges will be handled between the two OVSDB implementations. The ovs-vsctl implementation uses sudo/rootwrap whereas the IDL library will require either giving the neutron user permissions to the ovsdb unix socket, or using TCP/SSL sockets and controlling access via firewall rules (though likely just using the loopback interface).
Any distributions or deployment targets that do not support the requirements for the proposed OVSDB implementation will still be able to use the existing implementation.
This will allow leaving the ovs_lib API largely unchanged. On a simple test creating and deleting 100 ports on an existing bridge, the current ovs_lib implementation was nearly 10x slower than using the OVS Python API. It should be noted that ovs_lib was over 100x slower than sending raw OVSDB commands to Open vSwitch. The performance disparity between the raw OVSDB and existing OVS Python API is due to the poor performance of the OVS Python API’s own pure-Python JSON parser which tests show to be ~30x slower than Python’s stdlib JSON parser which, unfortunately, is not a good fit for OVS’s use case as it requires having the entire string ready to parse, whereas the OVS code is structured to parse buffers as they are read. Potentially significant speedups are possible in OVS’s JSON parser by writing an extension that adds Python bindings to the C version of their parser.
The OVS IDL implementation makes use of monitor commands for syncing a local cache of the OVSDB. Unfortunately, it does not make these monitor events available to users of the library. To replace ovsdb-monitor, it will be necessary to either use a lower-level Open vSwitch API running an additional monitor request to get these event notifications, to modify the IDL library at runtime, or to try to get code merged to the upstream Open vSwitch library that optionally exposes these events.
The existing implementation handles privileges via sudo/rootwrap. The proposed change would be using a programatic API and would instead rely on the appropriate permissions being set on underlying Open vSwitch unix domain socket or through the use of Open vSwitch’s SSL authentication.
With no changes to the upstream OVS Python API, a 6-7x speed improvement is possible. With Python bindings to the OVS C JSON parser, it should be possible to approach the native OVSDB protocol performance which was 100x faster than the current ovs_lib.
Repeating from the Security section, deployers would have to ensure that the neutron user has r/w permissions to the Open vSwitch db.sock. A new config option specifying the connection string for ovs-server would also be required. Packagers will need to add a dependency for python-openvswitch (which seems to be readily available across common distributions as it is part of standard openvswitch packaging).
For the initial implementation, there shouldn’t be much change to ovs_lib’s API.
It should be noted that users of XenAPI currently use a modified version of rootwrap to execute ovs-vsctl commands. To use the proposed OVSDB implementation with the XenAPI, an equivalent wrapper may have to be developed. This work is outside of the scope of this blueprint.
It would be possible to write our own library around the OVSDB protocol. It would most likely be faster, but it would most likely end up looking very much like the OVS Python API by the time we were finished.
Adds a requirement for the OVS python bindings
Many of the unit tests for ovs_lib will have to be changed because they are tied far too heavily to specific implementation details, mocking out calls to ovs-vsctl, etc.
Functional tests that actually test the CRUD operations against a real Open vSwitch installation should be created. They should work against both OVSDB interface implementations.
Documentation of the new config option and security considerations will be necessary.
The OVSDB abstract base class should be well-documented. In-tree developer docs describing the ovs_lib implementation will be added.