1 * Flow match expression handling library.
3 ovn-controller is the primary user of flow match expressions, but
4 the same syntax and I imagine the same code ought to be useful in
5 ovn-northd for ACL match expressions.
9 ** Flow table handling in ovn-controller.
11 ovn-controller has to transform logical datapath flows from the
12 database into OpenFlow flows.
14 *** Definition (or choice) of data structure for flows and flow table.
16 It would be natural enough to use "struct flow" and "struct
17 classifier" for this. Maybe that is what we should do. However,
18 "struct classifier" is optimized for searches based on packet
19 headers, whereas all we care about here can be implemented with a
20 hash table. Also, we may want to make it easy to add and remove
21 support for fields without recompiling, which is not possible with
22 "struct flow" or "struct classifier".
24 On the other hand, we may find that it is difficult to decide that
25 two OXM flow matches are identical (to normalize them) without a
26 lot of domain-specific knowledge that is already embedded in struct
27 flow. It's also going to be a pain to come up with a way to make
28 anything other than "struct flow" work with the ofputil_*()
29 functions for encoding and decoding OpenFlow.
31 It's also possible we could use struct flow without struct
34 *** Assembling conjunctive flows from flow match expressions.
36 This transformation explodes logical datapath flows into multiple
37 OpenFlow flow table entries, since a flow match expression in CoD
38 form requires several OpenFlow flow table entries. It also
39 requires merging together OpenFlow flow tables entries that contain
40 "conjunction" actions (really just concatenating their actions).
42 *** Translating logical datapath port names into port numbers.
44 Logical ports are specified by name in logical datapath flows, but
45 OpenFlow only works in terms of numbers.
47 *** Translating logical datapath actions into OpenFlow actions.
49 Some of the logical datapath actions do not have natural
50 representations as OpenFlow actions: they require
51 packet-in/packet-out round trips through ovn-controller. The
52 trickiest part of that is going to be making sure that the
53 packet-out resumes the control flow that was broken off by the
54 packet-in. That's tricky; we'll probably have to restrict control
55 flow or add OVS features to make resuming in general possible. Not
56 sure which is better at this point.
58 *** OpenFlow flow table synchronization.
60 The internal representation of the OpenFlow flow table has to be
61 synced across the controller connection to OVS. This probably
62 boils down to the "flow monitoring" feature of OF1.4 which was then
63 made available as a "standard extension" to OF1.3. (OVS hasn't
64 implemented this for OF1.4 yet, but the feature is based on a OVS
65 extension to OF1.0, so it should be straightforward to add it.)
67 We probably need some way to catch cases where OVS and OVN don't
68 see eye-to-eye on what exactly constitutes a flow, so that OVN
69 doesn't waste a lot of CPU time hammering at OVS trying to install
70 something that it's not going to do.
72 *** Logical/physical translation.
74 When a packet comes into the integration bridge, the first stage of
75 processing needs to translate it from a physical to a logical
76 context. When a packet leaves the integration bridge, the final
77 stage of processing needs to translate it back into a physical
78 context. ovn-controller needs to populate the OpenFlow flows
79 tables to do these translations.
81 *** Determine how to split logical pipeline across physical nodes.
83 From the original OVN architecture document:
85 The pipeline processing is split between the ingress and egress
86 transport nodes. In particular, the logical egress processing may
87 occur at either hypervisor. Processing the logical egress on the
88 ingress hypervisor requires more state about the egress vif's
89 policies, but reduces traffic on the wire that would eventually be
90 dropped. Whereas, processing on the egress hypervisor can reduce
91 broadcast traffic on the wire by doing local replication. We
92 initially plan to process logical egress on the egress hypervisor
93 so that less state needs to be replicated. However, we may change
94 this behavior once we gain some experience writing the logical
97 The split pipeline processing split will influence how tunnel keys
100 ** Interaction with Open_vSwitch and OVN databases:
102 *** Monitor Chassis table in OVN.
104 Populate Port records for tunnels to other chassis into
105 Open_vSwitch database. As a scale optimization later on, one can
106 populate only records for tunnels to other chassis that have
107 logical networks in common with this one.
109 *** Monitor Pipeline table in OVN, trigger flow table recomputation on change.
111 ** ovn-controller parameters and configuration.
113 *** Tunnel encapsulation to publish.
115 Default: VXLAN? Geneve?
117 *** SSL configuration.
119 Can probably get this from Open_vSwitch database.
123 ** Monitor OVN_Northbound database, trigger Pipeline recomputation on change.
125 ** Translate each OVN_Northbound entity into Pipeline logical datapath flows.
127 We have to first sit down and figure out what the general
128 translation of each entity is. The original OVN architecture
130 http://openvswitch.org/pipermail/dev/2015-January/050380.html had
131 some sketches of these, but they need to be completed and
134 Initially, the simplest way to do this is probably to write
135 straight C code to do a full translation of the entire
136 OVN_Northbound database into the format for the Pipeline table in
137 the OVN Southbound database. As scale increases, this will probably
138 be too inefficient since a small change in OVN_Northbound requires a
139 full recomputation. At that point, we probably want to adopt a more
140 systematic approach, such as something akin to the "nlog" system used
141 in NVP (see Koponen et al. "Network Virtualization in Multi-tenant
142 Datacenters", NSDI 2014).
144 ** Push logical datapath flows to Pipeline table.
146 ** Monitor OVN Southbound database Bindings table.
148 Sync rows in the OVN Bindings table to the "up" column in the
149 OVN_Northbound database.
153 ovsdb-server should have adequate features for OVN but it probably
154 needs work for scale and possibly for availability as deployments
155 grow. Here are some thoughts.
157 Andy Zhou is looking at these issues.
159 ** Scaling number of connections.
161 In typical use today a given ovsdb-server has only a single-digit
162 number of simultaneous connections. The OVN Southbound database will
163 have a connection from every hypervisor. This use case needs testing
164 and probably coding work. Here are some possible improvements.
166 *** Reducing amount of data sent to clients.
168 Currently, whenever a row monitored by a client changes,
169 ovsdb-server sends the client every monitored column in the row,
170 even if only one column changes. It might be valuable to reduce
171 this only to the columns that changes.
173 Also, whenever a column changes, ovsdb-server sends the entire
174 contents of the column. It might be valuable, for columns that
175 are sets or maps, to send only added or removed values or
178 Currently, clients monitor the entire contents of a table. It
179 might make sense to allow clients to monitor only rows that
180 satisfy specific criteria, e.g. to allow an ovn-controller to
181 receive only Pipeline rows for logical networks on its hypervisor.
183 *** Reducing redundant data and code within ovsdb-server.
185 Currently, ovsdb-server separately composes database update
186 information to send to each of its clients. This is fine for a
187 small number of clients, but it wastes time and memory when
188 hundreds of clients all want the same updates (as will be in the
191 (This is somewhat opposed to the idea of letting a client monitor
192 only some rows in a table, since that would increase the diversity
197 If it turns out that other changes don't let ovsdb-server scale
198 adequately, we can multithread ovsdb-server. Initially one might
199 only break protocol handling into separate threads, leaving the
200 actual database work serialized through a lock.
202 ** Increasing availability.
204 Database availability might become an issue. The OVN system
205 shouldn't grind to a halt if the database becomes unavailable, but
206 it would become impossible to bring VIFs up or down, etc.
208 My current thought on how to increase availability is to add
209 clustering to ovsdb-server, probably via the Raft consensus
210 algorithm. As an experiment, I wrote an implementation of Raft
211 for Open vSwitch that you can clone from:
213 https://github.com/blp/ovs-reviews.git raft
215 ** Reducing startup time.
217 As-is, if ovsdb-server restarts, every client will fetch a fresh
218 copy of the part of the database that it cares about. With
219 hundreds of clients, this could cause heavy CPU load on
220 ovsdb-server and use excessive network bandwidth. It would be
221 better to allow incremental updates even across connection loss.
222 One way might be to use "Difference Digests" as described in
223 Epstein et al., "What's the Difference? Efficient Set
224 Reconciliation Without Prior Context". (I'm not yet aware of
225 previous non-academic use of this technique.)
229 ** Write ovn-nbctl utility.
231 The idea here is that we need a utility to act on the OVN_Northbound
232 database in a way similar to a CMS, so that we can do some testing
233 without an actual CMS in the picture.
237 ** Init scripts for ovn-controller (on HVs), ovn-northd, OVN DB server.
239 ** Distribution packaging.
245 This is being developed on OpenStack's development infrastructure
246 to be along side most of the other Neutron plugins.
248 http://git.openstack.org/cgit/stackforge/networking-ovn
250 http://git.openstack.org/cgit/stackforge/networking-ovn/tree/doc/source/todo.rst