[openstack-dev] Fwd: FW: [Neutron] Group Based Policy and the way forward
Wuhongning
wuhongning at huawei.com
Fri Aug 8 14:22:21 UTC 2014
Does it make sense to move all advanced extension out of ML2, like security group, qos...? Then we can just talk about advanced service itself, without bothering basic neutron object (network/subnet/port)
Traditionally, SG is applied in CN, and FWaas is applied in NN (bound to L3 agent), however in DVR they are bluing, for each host has a L3 agent, this brings opportunity to move all service related features out of L2 agent, and coordinate the two security modules (SG & FW for W-E traffic). However, neutron agents (l2/l3/advanced service) needs some rework.
When all these service features is detached from ML2, we can easily keep the mainstream OVS continually as L2 backend, and vendor specific hardware as the service policy enforcement backend such as acl/qos for high performance.
So maybe a trade-off is better: an optionally new policy service plugin together with SG and FWaas, cloud operate can choose what to present to enduser, but without concept of EP/EPG/BD/RD(now renamed to L2Pand L3P), Policy only focus on service layer, and policy template will be applied to existing neutron port object.
EP/EPG/L2P/L3P is really not so friendly to endusers facing. I asked some SMB IT staff and personal user of amazon AWS (they all have very limited networking knowledge), they can easily understand port/network/subnet, but can't understand those GBP objects. So, maybe apply advanced service policy to existing basic neutron object is more smoothly for endusers to accept.
________________________________
From: Kevin Benton [blak111 at gmail.com]
Sent: Friday, August 08, 2014 2:42 PM
To: OpenStack Development Mailing List (not for usage questions)
Subject: Re: [openstack-dev] Fwd: FW: [Neutron] Group Based Policy and the way forward
Can you link to the etherpad you mentioned?
In the mean time, apologies for another analogy in
advance. :-)
If I give you an API to sort a list, I'm free to implement it however I want as long as I return a sorted list. However, there is no way me to know based on a call to this API that you might only be looking for the second largest element, so it won't be the most efficient approach because I will always have to sort the entire list.
If I give you a higher level API to declare that you want elements of a list that match a criteria in a certain order, then the API can make the optimization to not actually sort the whole list if you just need the first of the largest two elements.
The former is analogous to the security groups API, and the latter to the GBP API.
On Aug 7, 2014 4:00 PM, "Aaron Rosen" <aaronorosen at gmail.com<mailto:aaronorosen at gmail.com>> wrote:
On Thu, Aug 7, 2014 at 12:08 PM, Kevin Benton <blak111 at gmail.com<mailto:blak111 at gmail.com>> wrote:
>I mean't 'side stepping' why GBP allows for the comment you made previous, "With the latter, a mapping driver could determine that communication between these two hosts can be prevented by using an ACL on a router or a switch, which doesn't violate the user's intent and buys a performance improvement and works with ports that don't support security groups.".
>Neutron's current API is a logical abstraction and enforcement can be done however one chooses to implement it. I'm really trying to understand at the network level why GBP allows for these optimizations and performance improvements you talked about.
You absolutely cannot enforce security groups on a firewall/router that sits at the boundary between networks. If you try, you are lying to the end-user because it's not enforced at the port level. The current neutron APIs force you to decide where things like that are implemented.
The current neutron API's are just logical abstractions. Where and how things are actually enforced are 100% an implementation detail of a vendors system. Anyways, moving the discussion to the etherpad...
The higher level abstractions give you the freedom to move the enforcement by allowing the expression of broad connectivity requirements.
>Why are you bringing up logging connections?
This was brought up as a feature proposal to FWaaS because this is a basic firewall feature missing from OpenStack. However, this does not preclude a FWaaS vendor from logging.
>Personally, I think one could easily write up a very short document probably less than one page with examples showing/exampling how the current neutron API works even without a much networking background.
The difficulty of the API for establishing basic connectivity isn't really the problem. It's when you have to compose a bunch of requirements and make sure nothing is violating auditing and connectivity constraints that it becomes a problem. We are arguing about the levels of abstraction. You could also write up a short document explaining to novice programmers how to use C to read and write database entries to an sqlite database, but that doesn't mean it's the best level of abstraction for what the users are trying to accomplish.
I'll let someone else explain the current GBP API because I'm not working on that. I'm just trying to convince you of the value of declarative network configuration.
On Thu, Aug 7, 2014 at 12:02 PM, Aaron Rosen <aaronorosen at gmail.com<mailto:aaronorosen at gmail.com>> wrote:
On Thu, Aug 7, 2014 at 9:54 AM, Kevin Benton <blak111 at gmail.com<mailto:blak111 at gmail.com>> wrote:
You said you had no idea what group based policy was buying us so I tried to illustrate what the difference between declarative and imperative network configuration looks like. That's the major selling point of GBP so I'm not sure how that's 'side stepping' any points. It removes the need for the user to pick between implementation details like security groups/FWaaS/ACLs.
I mean't 'side stepping' why GBP allows for the comment you made previous, "With the latter, a mapping driver could determine that communication between these two hosts can be prevented by using an ACL on a router or a switch, which doesn't violate the user's intent and buys a performance improvement and works with ports that don't support security groups.".
Neutron's current API is a logical abstraction and enforcement can be done however one chooses to implement it. I'm really trying to understand at the network level why GBP allows for these optimizations and performance improvements you talked about.
>So are you saying that GBP allows someone to be able to configure an application that at the end of the day is equivalent to networks/router/FWaaS rules without understanding networking concepts?
It's one thing to understand the ports an application leverages and another to understand the differences between configuring VM firewalls, security groups, FWaaS, and router ACLs.
Sure, but how does group based policy solve this. Security Groups and FWaaS are just different places of enforcement. Say I want different security enforcement on my router than on my instances. One still needs to know enough to tell group based policy this right? They need to know enough that there are different enforcement points? How is doing this with Group based policy make it easier?
> I'm also curious how this GBP is really less error prone than the model we have today as it seems the user will basically have to tell neutron the same information about how he wants his networking to function.
With GBP, the user just gives the desired end result (e.g. allow connectivity between endpoint groups via TCP port 22 with all connections logged). Without it, the user has to do the following:
Why are you bringing up logging connections? Neutron has no concept of this at all today in it's code base. Is logging something related to GBP?
1. create a network/subnet for each endpoint group
2. allow all traffic on the security groups since the logging would need to be accomplished with FWaaS
3. create an FWaaS instance
4. attach the FWaaS to both networks
Today FWaaS api is still incomplete as there is no real point of enforcement in it's api (though it really seems that it should just be router ports) it's just global on the router right now.
1. add an FWaaS policy and the FWaaS rules to allow the correct traffic
I'd more or less agree with these steps. Would you mind also giving the steps involved in group based policy so we can compare?
The declarative approach is less error prone because the user can give neutron the desired state of connectivity rather than a mentally compiled set of instructions describing how to configure a bunch of individual network components. How well do you think someone will handle the latter approach that got all of their networking knowledge from one college course 5 years ago?
Personally, I think one could easily write up a very short document probably less than one page with examples showing/exampling how the current neutron API works even without a much networking background. The funny thing is I've been trying to completely understand the proposed group policy api for a little while now and I'm still having trouble. Seems like we're taking abstractions that are quite well known/understood and changing them to a different model that requires one to know about this new terminology:
Endpoint (EP): An L2/L3 addressable entity.
Endpoint Group (EPG): A collection of endpoints.
Contract: It defines how the application services provided by an EPG can be accessed. In effect it specifies how an EPG communicates with other EPGs. A Contract consists of Policy Rules.
Policy Rule: These are individual rules used to define the communication criteria between EPGs. Each rule contains a Filter, Classifier, and Action.
Classifier: Characterizes the traffic that a particular Policy Rule acts on. Corresponding action is taken on traffic that satisfies this classification criteria.
Action: The action that is taken for a matching Policy Rule defined in a Contract.
Filter: Provides a way to tag a Policy Rule with Capability and Role labels.
Capability: It is a Policy Label that defines what part of a Contract a particular EPG provides.
Role: It is a Policy Label that defines what part of a Contract an EPG wants to consume.
Contract Scope: An EPG conveys its intent to provide or consume a Contract (or its part) by defining a Contract Scope which references the target Contract.
Selector: A Contract Scope can define additional constraints around choosing the matching provider or consumer EPGs for a Contract via a Selector.
Policy Labels: These are labels contained within a namespace hierarchy and used to define Capability and Role tags used in Filters.
Bridge Domain: Used to define a L2 boundary and impose additional constraints (such as no broadcast) within that L2 boundary.
Routing Domain: Used to define a non-overlapping IP address space.
IIRC one of the early nova parity requests was an API to automagically setup a neutron network and router so no extra work would be required to get instances connected to the Internet. That wasn't requested because people thought neutron networks were too easy to setup already. :-)
I think the confusion why that comment was made is probably because the nova-networks model doesn't have a concept of routers. Neutron can operate in the same way though with flat and provider networks. Either way one could easily write a nova-binding to the command they used previously to create a router and uplink it hiding the fact that there is a router there.
On Thu, Aug 7, 2014 at 9:10 AM, Aaron Rosen <aaronorosen at gmail.com<mailto:aaronorosen at gmail.com>> wrote:
Hi Kevin,
I feel as your latest response is completely side stepping the points we have been trying to get to in the last series of emails. At the end of the day I don't believe we are changing the laws of networking (or perhaps we are?). Thus I think it's important to actually get down to the networking level to actually figure out why optimizations such as this one are enabled via GBP and not the model we have:
"With the latter, a mapping driver could determine that communication between these two hosts can be prevented by using an ACL on a router or a switch, which doesn't violate the user's intent and buys a performance improvement and works with ports that don't support security groups."
On Wed, Aug 6, 2014 at 7:48 PM, Kevin Benton <blak111 at gmail.com<mailto:blak111 at gmail.com>> wrote:
Do you not see a difference between explicitly configuring networks, a router and FWaaS rules with logging and just stating that two groups of servers can only communicate via one TCP port with all connections logged? The first is very prone to errors for someone deploying an application without a strong networking background, and the latter is basically just stating the requirements and letting Neutron figure out how to implement it.
So are you saying that GBP allows someone to be able to configure an application that at the end of the day is equivalent to networks/router/FWaaS rules without understanding networking concepts? I'm also curious how this GBP is really less error prone than the model we have today as it seems the user will basically have to tell neutron the same information about how he wants his networking to function.
Just stating requirements becomes even more important when something like the logging requirement comes from someone other than the app deployer (e.g. a security team). In the above example, someone could set everything up using security groups; however, when the logging requirement came in from the security team, they would have to undo all of that work and replace it with something like FWaaS that can centrally log all of the connections.
It's the difference between using puppet and bash scripts. Sure you can write a script that uses awk/sed to ensure that an ini file has a particular setting and then restart a service if the setting is changed, but it's much easier and less error prone to just write a puppet manifest that uses the INI module with a pointer to the file, the section name, the key, and the value with a notification to restart the service.
On Wed, Aug 6, 2014 at 7:40 PM, Aaron Rosen <aaronorosen at gmail.com<mailto:aaronorosen at gmail.com>> wrote:
On Wed, Aug 6, 2014 at 5:27 PM, Kevin Benton <blak111 at gmail.com<mailto:blak111 at gmail.com>> wrote:
Web tier can communicate with anything except for the DB.
App tier can only communicate with Web and DB.
DB can communicate with App.
These high-level constraints can then be implemented as security groups like you showed, or network hardware ACLs like I had shown.
But if you start with the security groups API, you are forcing it to be implemented there.
I still have no idea what group based policy is buying us then. It seems to me that the key point we've identified going backing and forth here is the difference between the current model and the GBP model is that GBP constricts topology which allows us to write these types of enforcement rules. The reason we want this is because it yields performance optimizations (for some reason, which I don't think we've gotten into). Would you agree this is accurate?
Honestly, I know a lot of work has been put into this. I haven't said I'm for or against it either. I'm really just trying to understand what is the motivation for this and why does it make neutron better.
Best,
Aaron
On Wed, Aug 6, 2014 at 6:06 PM, Aaron Rosen <aaronorosen at gmail.com<mailto:aaronorosen at gmail.com>> wrote:
On Wed, Aug 6, 2014 at 4:46 PM, Kevin Benton <blak111 at gmail.com<mailto:blak111 at gmail.com>> wrote:
That's the point. By using security groups, you are forcing a certain kind of enforcement that must be honored and might not be necessary if the original intent was just to isolate between groups. In the example you gave, it cannot be implemented on the router without violating the constraints of the security group.
Hi Kevin,
Mind proposing an alternative example then. The only way I can see this claim to be made is because Group Based policy is actually limiting what a tenants desired topology can be?
Thanks,
Aaron
On Wed, Aug 6, 2014 at 5:39 PM, Aaron Rosen <aaronorosen at gmail.com<mailto:aaronorosen at gmail.com>> wrote:
On Wed, Aug 6, 2014 at 4:18 PM, Kevin Benton <blak111 at gmail.com<mailto:blak111 at gmail.com>> wrote:
>Given this information I don't see any reason why the backend system couldn't do enforcement at the logical router and if it did so neither parties would know.
With security groups you are specifying that nothing can contact these devices on those ports unless they come from the allowed IP addresses. If you tried to enforce this at the router you would be violating that specification because devices in the same subnet would be able to communicate on those blocked ports.
Sure, though this is a problem of where you are doing your enforcement. If the backend system chooses to implement this optimization in this fashion (which was the example you gave previously [1]). Then, if the topology changes, i.e adding a port to the same network with conflicting security group rules, the backend system can no longer optimize in this same fashion at the router level and a more fine grain filtering will need to be done. How would this be any different with group based policy?
[1] - With the latter, a mapping driver could determine that communication between these two hosts can be prevented by using an ACL on a router or a switch, which doesn't violate the user's intent and buys a performance improvement and works with ports that don't support security groups.
states
On Wed, Aug 6, 2014 at 5:00 PM, Aaron Rosen <aaronorosen at gmail.com<mailto:aaronorosen at gmail.com>> wrote:
On Wed, Aug 6, 2014 at 3:35 PM, Kevin Benton <blak111 at gmail.com<mailto:blak111 at gmail.com>> wrote:
By working at the port level you have already eliminated your ability to implement the filtering at different components of the network. They now need to be implemented in stateful rules at the port level and the device has to support security groups.
Lets take this example where we setup a 2 tier app with web-servers and db-servers that are connected on two different networks attached to a router. We add a security group rules such that web-servers can talk to db-servers on tcp:3306 and a rule to allow tcp:80 into the web-servers from anywhere.
neutron net-create web_net
neutron subnet-create --name web_subnet web_net 10.0.0.0/24<http://10.0.0.0/24>
neutron net-create db_net
neutron subnet-create --name db_subnet db_net 10.2.0.0/24<http://10.2.0.0/24>
neutron router-create myrouter
neutron router-interface-add myrouter web_subnet
neutron router-interface-add myrouter db_subnet
neutron security-group-create web-servers;
neutron security-group-create db-servers;
# add rule to allow web members to talk to the db-servers on TCP 3306 for their db connection;
neutron security-group-rule-create --protocol TCP --port-range-min 3306 --port-range-max 3306 --remote-group-id web-servers db-servers
# add rule to allow TCP 80 into the web-server sg
neutron security-group-rule-create --protocol TCP --port-range-min 80 --port-range-max 80 web-servers db-servers
# create some ports with desired security profiles.
neutron port-create --security-group web-servers web_net
neutron port-create --security-group web-servers web_net
neutron port-create --security-group db-servers db_net
neutron port-create --security-group db-servers db_net
Now to your point:
By working at the port level you have already eliminated your ability to implement the filtering at different components of the network. They now need to be implemented in stateful rules at the port level and the device has to support security groups.
Given this information I don't see any reason why the backend system couldn't do enforcement at the logical router and if it did so neither parties would know. The backend system should have the full graph of everything and be able to do enforcement optimizations where ever it likes.
btw: I say the enforcement could be done on the logical router though the backend system could also do this on the physical fabic as well if it wanted to as it should also know that graph. No?
On Wed, Aug 6, 2014 at 4:03 PM, Aaron Rosen <aaronorosen at gmail.com<mailto:aaronorosen at gmail.com>> wrote:
On Wed, Aug 6, 2014 at 12:46 PM, Kevin Benton <blak111 at gmail.com<mailto:blak111 at gmail.com>> wrote:
>I believe the referential security group rules solve this problem (unless I'm not understanding):
I think the disconnect is that you are comparing the way to current mapping driver implements things for the reference implementation with the existing APIs. Under this light, it's not going to look like there is a point to this code being in Neutron since, as you said, the abstraction could happen at a client. However, this changes once new mapping drivers can be added that implement things differently.
Let's take the security groups example. Using the security groups API directly is imperative ("put a firewall rule on this port that blocks this IP") compared to a higher level declarative abstraction ("make sure these two endpoints cannot communicate"). With the former, the ports must support security groups and there is nowhere except for the firewall rules on that port to implement it without violating the user's expectation. With the latter, a mapping driver could determine that communication between these two hosts can be prevented by using an ACL on a router or a switch, which doesn't violate the user's intent and buys a performance improvement and works with ports that don't support security groups.
Group based policy is trying to move the requests into the declarative abstraction so optimizations like the one above can be made.
Hi Kevin,
Interesting points. Though, let me ask this. Why do we need to move to a declarative API abstraction in neutron in order to perform this optimization on the backend? For example, In the current neutron model say we want to create a port with a security group attached to it called web that allows TCP:80 in and members who are in a security group called database. From this mapping I fail to see how it's really any different from the declarative model? The ports in neutron are logical abstractions and the backend system could be implemented in order to determine that the communication between these two hosts could be prevented by using an ACL on a router or switch as well.
Best,
Aaron
_______________________________________________
OpenStack-dev mailing list
OpenStack-dev at lists.openstack.org<mailto:OpenStack-dev at lists.openstack.org>
http://lists.openstack.org/cgi-bin/mailman/listinfo/openstack-dev
_______________________________________________
OpenStack-dev mailing list
OpenStack-dev at lists.openstack.org<mailto:OpenStack-dev at lists.openstack.org>
http://lists.openstack.org/cgi-bin/mailman/listinfo/openstack-dev
--
Kevin Benton
_______________________________________________
OpenStack-dev mailing list
OpenStack-dev at lists.openstack.org<mailto:OpenStack-dev at lists.openstack.org>
http://lists.openstack.org/cgi-bin/mailman/listinfo/openstack-dev
_______________________________________________
OpenStack-dev mailing list
OpenStack-dev at lists.openstack.org<mailto:OpenStack-dev at lists.openstack.org>
http://lists.openstack.org/cgi-bin/mailman/listinfo/openstack-dev
--
Kevin Benton
_______________________________________________
OpenStack-dev mailing list
OpenStack-dev at lists.openstack.org<mailto:OpenStack-dev at lists.openstack.org>
http://lists.openstack.org/cgi-bin/mailman/listinfo/openstack-dev
_______________________________________________
OpenStack-dev mailing list
OpenStack-dev at lists.openstack.org<mailto:OpenStack-dev at lists.openstack.org>
http://lists.openstack.org/cgi-bin/mailman/listinfo/openstack-dev
--
Kevin Benton
_______________________________________________
OpenStack-dev mailing list
OpenStack-dev at lists.openstack.org<mailto:OpenStack-dev at lists.openstack.org>
http://lists.openstack.org/cgi-bin/mailman/listinfo/openstack-dev
_______________________________________________
OpenStack-dev mailing list
OpenStack-dev at lists.openstack.org<mailto:OpenStack-dev at lists.openstack.org>
http://lists.openstack.org/cgi-bin/mailman/listinfo/openstack-dev
--
Kevin Benton
_______________________________________________
OpenStack-dev mailing list
OpenStack-dev at lists.openstack.org<mailto:OpenStack-dev at lists.openstack.org>
http://lists.openstack.org/cgi-bin/mailman/listinfo/openstack-dev
_______________________________________________
OpenStack-dev mailing list
OpenStack-dev at lists.openstack.org<mailto:OpenStack-dev at lists.openstack.org>
http://lists.openstack.org/cgi-bin/mailman/listinfo/openstack-dev
--
Kevin Benton
_______________________________________________
OpenStack-dev mailing list
OpenStack-dev at lists.openstack.org<mailto:OpenStack-dev at lists.openstack.org>
http://lists.openstack.org/cgi-bin/mailman/listinfo/openstack-dev
_______________________________________________
OpenStack-dev mailing list
OpenStack-dev at lists.openstack.org<mailto:OpenStack-dev at lists.openstack.org>
http://lists.openstack.org/cgi-bin/mailman/listinfo/openstack-dev
--
Kevin Benton
_______________________________________________
OpenStack-dev mailing list
OpenStack-dev at lists.openstack.org<mailto:OpenStack-dev at lists.openstack.org>
http://lists.openstack.org/cgi-bin/mailman/listinfo/openstack-dev
_______________________________________________
OpenStack-dev mailing list
OpenStack-dev at lists.openstack.org<mailto:OpenStack-dev at lists.openstack.org>
http://lists.openstack.org/cgi-bin/mailman/listinfo/openstack-dev
--
Kevin Benton
_______________________________________________
OpenStack-dev mailing list
OpenStack-dev at lists.openstack.org<mailto:OpenStack-dev at lists.openstack.org>
http://lists.openstack.org/cgi-bin/mailman/listinfo/openstack-dev
_______________________________________________
OpenStack-dev mailing list
OpenStack-dev at lists.openstack.org<mailto: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/20140808/91863ba7/attachment-0001.html>
More information about the OpenStack-dev
mailing list