Dragonflow control plane test

New Dragonflow version has asynchronous architecture. On each object creation command, happends the following:

In Neutron: 1. Neutron receives an API call to create an object (network, subnet, port, etc...) 2. Neutron calls Dragonflow plugin with just created object 3. Dragonflow plugin saves a new record to NoSQL database 4. Dragonflow Plugin sends a message with the new object id using the MQ Pub/Sub

In Dragonflow controller 1. DF controller receives a message from MQ Pub/Sub with an object id 2. DF controller fetches full object records from NoSQL db 3. DF controller, if required, creates nessesary OpenFlow rules.

Now a new object is basically transfered to the actuall OpenFlow rules.

In the control plane test we will basically test all components of the Dragonflow architecture.

This test plan consists of the following sections: 1. NoSQL Database Test Plan 2. Message Publish Subscribe Test Plan 3. Neutron Objects Set-Up Time 4. Stress Tests

Out of scope:

1. Evaluation of network throughput in case the system has hundreds of security rules. We check only time required to deploy these rules in this project.

Our goals:

In this project we would like to see that the time required to add new configuration element to the system will not exceed XXX Miliseconds and we do not have a noticable degradation when the system has more configuration entities.

1. NoSQL Database Test Plan

Dragonflow can work with a number of NoSQL databases. Each database has t’s own advantages and disadvantages.

At the time of writing this document, Dragonflow supports the following key / value databases:

  1. Etcd
  2. Redis
  3. Ramcloud
  4. Zookeeper

Some of the databases above support clustering. We will perform tests against database server configured in single and multiple node.

The system will be tested against the following configuration:

  1. All in one server together with NoSQL server
  2. Two or more servers required to enable NoSQL clustering configuration.

Some of the results will be compared with MySQL. See bellow.

NoSQL random access test scenario

The test will create 1000 database objects and performs random access test. We can optionally compare the results agains MySQL by reading values from Neutron DB. These tests will mimic the case when Dragonflow receives notification when an object is created and it needs to load object records from the database.

1. Preparation

The script will create 1000 database objects like routers.

TODO: Create a script here

2. Random access script

The test will fetch a list of all objects from the database and fetch object in random way. The script will be tested against all supported databases.

TODO: create a script here

3. Record DB process activity

TODO: Create a script that records db process activity

3. Generate report

Generate comparison table.

2. Message Publish Subscribe Test Plan

Message publish/subscribe test plan is intended to test Dragonflow PUB/SUB mechanizm. This test measures the aggregate throughput of a MQ layer.

This test will cover the following: 1. ZeroMQ Pub/Sub 2. Redis Pub/Sub 3. Polling for DB changes

3. Neutron Objects Set-Up Time

New Dragonflow version has asynchronous architecture. As covered before it has 2 main parts: 1. Neutron API call handling with Dragonflow Plugin and pushing a notification to messaging queue 2. Actual processing on the API call in Dragonflow controller

In our test we will tests the following: 1. Test time for Neutron to handle API call and push notification to messaging queue 2. The same as above with actual time required to handle object creation/modification request in Dragonflow controller.

We will test the following Neutron configuration entities

  1. Network
  2. Subnetwork
  3. Router
  4. Security rules
  5. Security groups
  6. Network ports
  7. Floating ips

Basic test at zero state

We will calculate time to create multiple objects when system is at zero state. We define zero state as a state where we have a system with default rules only. We will do the following tests:

  1. 1 object created/updated
  2. 5 objects created/updated
  3. 10 objects created/updated
  4. 20 objects created/updated
  5. 50 objects created/updated
  6. 100 objects created/updated
  7. 200 objects created/updated
  8. 300 objects created/updated

9. 400 objects created/updated 8. 500 objects created/updated 9. 600 objects created/updated 10. 700 objects created/updated 10. 800 objects created/updated 10. 900 objects created/updated 10. 1000 objects created/updated

Multiple tenants

As Dragonflow addresses different tenants as different pub/sub channels, notification on object created in forwarded to one of another channel. So, the results, when using one tenant and multiple tennants will be different.

Heavy usage of the system (at single box)

In this test we will pre-create a lot of objects in the system and then we will measure time to add a new object to the system that is actively used.

What we are going to test

1. Check that objects are created are valid and correct Openflow rules are created 1. We will measure time to create one or group of objects 2. We will measure CPU usage

Now we will be able to perfom regression tests and compare results with new and old Dragonflow versions. In addition, we can run similar tests against the Neutron OVN deployment and compare results with the Neutron Dragonflow deployment.

4. Stress Tests Scenarios

In this test we want to stretch the system to it’s maximal capabilities and calculate time required in different scenarios.

For example we want to see how many small VM’s we can launch on a single box and how much time it takes to deploy all of them. In addition, we want to check that all of the VMs got an IP address.

Test scenarios for single server installation: 1. 1000 updates on one subnet (enable / disable DHCP) 1. 1 Router with 1000 Subnetworks 2. 1000 Routers - 1000 Subnetwork (1 subnetwork in 1 router) 3. 100 Routers - 500 subnets (5 subnets per router) 4. 1000 Security rules for 1 VM 5. 1000 Security rules for 10 VMs 6. Launch 200 VMs 7. Set up 1000 Security rules in 1 Security group 8. Etc...

Additional action items:

“There is also a control plane performance issue when we try to catch on the spec of typical AWS limit (200 subnets per router). When a router with 200 subnets is scheduled on a new host, a 30s delay is watched when all data plane setup is finished.”

“Create max Subnet on a router or for a tenant test create 1000 SG etc”