I spent a lot of my day assembling the following detailed to-do list
for OVN. It's in Emacs' outline-mode format, that is, the number of
"*"s indicates a nesting level.
I'm mostly posting this as a status report for people interested in
OVN. (But if anyone really wants to volunteer for some piece, let's
talk. I feel like I have dibs on some bits but there's a lot to do.)
I'm sure I missed a lot, please do send extra bullet points my way.
I'm adding this to the OVN RFC patch as ovn/TODO. There's not much
point in reposting the patch since nothing else has changed. Still,
if you really want, you can grab the revised commit from
https://github.com/blp/ovs-reviews.git ovn
----------------------------------------------------------------------
* Flow match expression handling library.
ovn-controller is the primary user of flow match expressions, but
the same syntax and I imagine the same code ought to be useful in
ovn-nbd for ACL match expressions.
** Definition of data structures to represent a match expression as a
syntax tree.
** Definition of data structures to represent variables (fields).
Fields need names and prerequisites. Most fields are numeric and
thus need widths. We need also need a way to represent nominal
fields (currently just logical port names). It might be
appropriate to associate fields directly with OXM/NXM code points;
we have to decide whether we want OVN to use the OVS flow structure
or work with OXM more directly.
Probably should be defined so that the data structure is also
useful for references to fields in action parsing.
** Lexical analysis.
Probably should be defined so that the lexer can be reused for
parsing actions.
** Parsing into syntax tree.
** Semantic checking against variable definitions.
** Applying prerequisites.
** Simplification into conjunction-of-disjunctions (CoD) form.
** Transformation from CoD form into OXM matches.
* ovn-controller
** Flow table handling in ovn-controller.
ovn-controller has to transform logical datapath flows from the
database into OpenFlow flows.
*** Definition (or choice) of data structure for flows and flow table.
It would be natural enough to use "struct flow" and "struct
classifier" for this. Maybe that is what we should do. However,
"struct classifier" is optimized for searches based on packet
headers, whereas all we care about here can be implemented with a
hash table. Also, we may want to make it easy to add and remove
support for fields without recompiling, which is not possible with
"struct flow" or "struct classifier".
On the other hand, we may find that it is difficult to decide that
two OXM flow matches are identical (to normalize them) without a
lot of domain-specific knowledge that is already embedded in struct
flow. It's also going to be a pain to come up with a way to make
anything other than "struct flow" work with the ofputil_*()
functions for encoding and decoding OpenFlow.
It's also possible we could use struct flow without struct
classifier.
*** Assembling conjunctive flows from flow match expressions.
This transformation explodes logical datapath flows into multiple
OpenFlow flow table entries, since a flow match expression in CoD
form requires several OpenFlow flow table entries. It also
requires merging together OpenFlow flow tables entries that contain
"conjunction" actions (really just concatenating their actions).
*** Translating logical datapath port names into port numbers.
Logical ports are specified by name in logical datapath flows, but
OpenFlow only works in terms of numbers.
*** Translating logical datapath actions into OpenFlow actions.
Some of the logical datapath actions do not have natural
representations as OpenFlow actions: they require
packet-in/packet-out round trips through ovn-controller. The
trickiest part of that is going to be making sure that the
packet-out resumes the control flow that was broken off by the
packet-in. That's tricky; we'll probably have to restrict control
flow or add OVS features to make resuming in general possible. Not
sure which is better at this point.
*** OpenFlow flow table synchronization.
The internal representation of the OpenFlow flow table has to be
synced across the controller connection to OVS. This probably
boils down to the "flow monitoring" feature of OF1.4 which was then
made available as a "standard extension" to OF1.3. (OVS hasn't
implemented this for OF1.4 yet, but the feature is based on a OVS
extension to OF1.0, so it should be straightforward to add it.)
We probably need some way to catch cases where OVS and OVN don't
see eye-to-eye on what exactly constitutes a flow, so that OVN
doesn't waste a lot of CPU time hammering at OVS trying to install
something that it's not going to do.
*** Logical/physical translation.
When a packet comes into the integration bridge, the first stage of
processing needs to translate it from a physical to a logical
context. When a packet leaves the integration bridge, the final
stage of processing needs to translate it back into a physical
context. ovn-controller needs to populate the OpenFlow flows
tables to do these translations.
*** Determine how to split logical pipeline across physical nodes.
From the original OVN architecture document:
The pipeline processing is split between the ingress and egress
transport nodes. In particular, the logical egress processing may
occur at either hypervisor. Processing the logical egress on the
ingress hypervisor requires more state about the egress vif's
policies, but reduces traffic on the wire that would eventually be
dropped. Whereas, processing on the egress hypervisor can reduce
broadcast traffic on the wire by doing local replication. We
initially plan to process logical egress on the egress hypervisor
so that less state needs to be replicated. However, we may change
this behavior once we gain some experience writing the logical
flows.
The split pipeline processing split will influence how tunnel keys
are encoded.
** Interaction with Open_vSwitch and OVN databases:
*** Monitor VIFs attached to the integration bridge in Open_vSwitch.
In response to changes, add or remove corresponding rows in
Bindings table in OVN.
*** Populate Chassis row in OVN at startup. Maintain Chassis row over time.
(Warn if any other Chassis claims the same IP address.)
*** Remove Chassis and Bindings rows from OVN on exit.
*** Monitor Chassis table in OVN.
Populate Port records for tunnels to other chassis into
Open_vSwitch database. As a scale optimization later on, one can
populate only records for tunnels to other chassis that have
logical networks in common with this one.
*** Monitor Pipeline table in OVN, trigger flow table recomputation on change.
** ovn-controller parameters and configuration.
*** Tunnel encapsulation to publish.
We can probably default to GRE. VXLAN is more modern but it only
has a 24-bit key. STT has a 64-bit key but it's not ubiquitously
available.
*** Location of Open_vSwitch database.
We can probably use the same default as ovs-vsctl.
*** Location of OVN database.
Probably no useful default.
*** SSL configuration.
Can probably get this from Open_vSwitch database.
* ovn-nbd
** Monitor OVN_Northbound database, trigger Pipeline recomputation on change.
** Translate each OVN_Northbound entity into Pipeline logical datapath flows.
We have to first sit down and figure out what the general
translation of each entity is. The original OVN architecture
description at
http://openvswitch.org/pipermail/dev/2015-January/050380.html had
some sketches of these, but they need to be completed and
elaborated.
Initially, the simplest way to do this is probably to write
straight C code to do a full translation of the entire
OVN_Northbound database into the format for the Pipeline table in
the OVN database. As scale increases, this will probably be too
inefficient since a small change in OVN_Northbound requires a full
recomputation. At that point, we probably want to adopt a more
systematic approach, such as something akin to the "nlog" system
used in NVP (see Koponen et al. "Network Virtualization in
Multi-tenant Datacenters", NSDI 2014).
** Push logical datapath flows to Pipeline table.
** Monitor OVN database Bindings table.
Sync rows in the OVN Bindings table to the "up" column in the
OVN_Northbound database.
* ovsdb-server
ovsdb-server should have adequate features for OVN but it probably
needs work for scale and possibly for availability as deployments
grow. Here are some thoughts.
Andy Zhou is looking at these issues.
** Scaling number of connections.
In typical use today a given ovsdb-server has only a single-digit
number of simultaneous connections. The OVN database will have a
connection from every hypervisor. This use case needs testing and
probably coding work. Here are some possible improvements.
*** Reducing amount of data sent to clients.
Currently, whenever a row monitored by a client changes,
ovsdb-server sends the client every monitored column in the row,
even if only one column changes. It might be valuable to reduce
this only to the columns that changes.
Also, whenever a column changes, ovsdb-server sends the entire
contents of the column. It might be valuable, for columns that
are sets or maps, to send only added or removed values or
key-values pairs.
Currently, clients monitor the entire contents of a table. It
might make sense to allow clients to monitor only rows that
satisfy specific criteria, e.g. to allow an ovn-controller to
receive only Pipeline rows for logical networks on its hypervisor.
*** Reducing redundant data and code within ovsdb-server.
Currently, ovsdb-server separately composes database update
information to send to each of its clients. This is fine for a
small number of clients, but it wastes time and memory when
hundreds of clients all want the same updates (as will be in the
case in OVN).
(This is somewhat opposed to the idea of letting a client monitor
only some rows in a table, since that would increase the diversity
among clients.)
*** Multithreading.
If it turns out that other changes don't let ovsdb-server scale
adequately, we can multithread ovsdb-server. Initially one might
only break protocol handling into separate threads, leaving the
actual database work serialized through a lock.
** Increasing availability.
Database availability might become an issue. The OVN system
shouldn't grind to a halt if the database becomes unavailable, but
it would become impossible to bring VIFs up or down, etc.
My current thought on how to increase availability is to add
clustering to ovsdb-server, probably via the Raft consensus
algorithm. As an experiment, I wrote an implementation of Raft
for Open vSwitch that you can clone from:
https://github.com/blp/ovs-reviews.git raft
** Reducing startup time.
As-is, if ovsdb-server restarts, every client will fetch a fresh
copy of the part of the database that it cares about. With
hundreds of clients, this could cause heavy CPU load on
ovsdb-server and use excessive network bandwidth. It would be
better to allow incremental updates even across connection loss.
One way might be to use "Difference Digests" as described in
Epstein et al., "What's the Difference? Efficient Set
Reconciliation Without Prior Context". (I'm not yet aware of
previous non-academic use of this technique.)
* Miscellaneous:
** Write ovn-nbctl utility.
The idea here is that we need a utility to act on the OVN_Northbound
database in a way similar to a CMS, so that we can do some testing
without an actual CMS in the picture.
No details yet.
** Init scripts for ovn-controller (on HVs), ovn-nbd, OVN DB server.
** Distribution packaging.
* Not yet scoped:
** Neutron plugin.
I'd appreciate input on this.
** Gateways.
_______________________________________________
dev mailing list
dev@openvswitch.org
http://openvswitch.org/mailman/listinfo/dev
