OpenNESS 19.12 Investigation

OpenNESS 19.12 Design

Openness released 19.12 on December 21 2019 and this new release has removed the deployment mode ( kubernetes + NTS ). Two modes is supported now: Native deployment Mode (which is based on pure docker/libvirt) and Infrastructure Mode (which is based on kube-ovn), below are the brief summary of the difference of these 2 modes:

Gap Analysis for Integrating OpenNESS with ICN

Network Policy

Network policy and DNS is used for traffic steering. Network policy is used for restrict access among services but NOT “proactively” forward the traffic, While the OpenNESS DNS service can help “redirect” the external client’s traffic to the edge application service。

By default, in a Network Edge environment, all ingress traffic is blocked (services running inside of deployed applications are not reachable) and all egress traffic is enabled (pods are able to reach the internet). The following NetworkPolicy definition is used:

apiVersion: networking.k8s.io/v1
metadata:
  name: block-all-ingress
  namespace: default        # selects default namespace
spec:
  podSelector: {}           # matches all the pods in the default namespace
  policyTypes:
  - Ingress
  ingress: []               # no rules allowing ingress traffic = ingress blocked

Admin can enable access to certain service by applying a NetworkPolicy CRD. For example:

1. To deploy a Network Policy allowing ingress traffic on port 5000 (tcp and udp) from 192.168.1.0/24 network to OpenVINO consumer application pod, create the following specification file for this Network Policy:

apiVersion: networking.k8s.io/v1
kind: NetworkPolicy
metadata:
  name: openvino-policy
  namespace: default
spec:
  podSelector:
    matchLabels:
      name: openvino-cons-app
  policyTypes:
  - Ingress
  ingress:
  - from:
    - ipBlock:
        cidr: 192.168.1.0/24
    ports:
    - protocol: TCP
      port: 5000
    - protocol: UDP
      port: 5000

2. Create the Network Policy:
kubectl apply -f network_policy.yml

DNS

DNS service can help “redirect” the external client’s traffic to the edge application service. This gap analysis is to investigate whether OpenNESS DNS can be used for ICN traffic steering or not.

OpenNESS provides DNS server which provides the microsevice’s ip address based on FQDN. OpenNESS extends kubectl utility with kubectl edgedns cmd to set/delete DNS entry. For example，

1. define a file with below content: openvino-dns.json
   {
     "record_type":"A",
     "fqdn":"openvino.openness",
     "addresses":["10.16.0.10"]
   }
2. Then use below command to add an entry in OpenNESS DNS server:
   kubectl edgedns set <edge_node_host_name> openvino-dns.json

Below are implement details of OpenNESS DNS server:

• Run as independent process/container in each Edge Node : ./edgednssvr -port 53 -fwdr=8.8.8.8 -db XXX.db // port: DNS server port; fwdr: forwarder ip used when cannot found FQDN in OpenNESS DNS DB; db: OpenNESS db file
• Provide 2 servers after running:
        Control Server: gRPC/IP based API to receive DNS record add/remove request – OpenNESS controller can call this interface to add DNS record
        DNS server: DNS service is based on https://github.com/miekg/dns
• DNS process flow: After get a DNS request, it will try to find the FQDN in local OpenNESS DNS db first, if not found, forward the request to an external forwarder (default is 8.8.8.8, set by “-fwdr“ parameter)

The OpenNESS DNS service is different from K8s’ CoreDNS to support different usages:

• CoreDNS: provides DNS service within K8s cluster, e.g. from app in container to find the service also running in container of the same cluster.
• OpenNESS DNS: provides DNS service for app of external host which is not running in the edge cluster to find a app (which may not be a K8s service, so its ip may not be recorded in coreDNS) in k8s cluster. e.g. in OpenNESS OpenVINO demo, the video stream generator is running in a separate host, admin needs manually (add a new name server in /etc/resolv.conf) set it’s DNS server IP to point to OpenNESS edge node DNS server then it can know how to send the stream.

Cross-Node communication

Edge apps can be divided into producer and consumer. This gap analysis is to investigate the communication between the producers and consumers which are on different edge nodes. 

Edge applications must introduce themselves to OpenNESS framework and identify if they would like to activate new edge services or consume an existing service. Edge Application Agent (EAA) component is the handler of all the edge applications hosted by the OpenNESS edge node and acts as their point-of-contact. 

OpenNESS-awareness involves (a) authentication, (b) service activation/deactivation, (c) service discovery, (d) service subscription, and (e) Websocket connection establishment. The Websocket connection retains a channel for EAA for notification forwarding to pre-subscribed consumer applications. Notifications are generated by "producer" edge applications and absorbed by "consumer" edge applications.

The sequence of operations for the producer application:

1. Authenticate with OpenNESS edge node
2. Activate new service and include the list of notifications involved
3. Send notifications to OpenNESS edge node according to business logic

The sequence of operations for the consumer application:

1. Authenticate with OpenNESS edge node
2. Discover the available services on OpenNESS edge platform
3. Subscribe to services of interest and listen for notifications

Edge apps will access eaa through eaa.openness (name.namespace) which is a kubernetes service:
https://github.com/open-ness/edgecontroller/blob/master/kube-ovn/openness.yaml#L18

For example: as following links show, openvino consumer will access http://eaa.openness:443/auth for authentication.
https://github.com/open-ness/edgeapps/blob/master/openvino/consumer/cmd/main.go#L24
https://github.com/open-ness/edgeapps/blob/master/openvino/consumer/cmd/main.go#L66

eaa is deployed as a deployment and only 1 eaa will be deployed:
https://github.com/open-ness/edgecontroller/blob/master/kube-ovn/openness.yaml#L41

Because all edge apps will access only 1 eaa, it doesn't matter that eaa is stateful. 

For example:
only 1 eaa is deployed on node1. producer1 and producer2 will activate the new service with eaa. consumer1 and consumer2 will consume services stored in eaa. Because all the information are stored in only 1 eaa, there won't be issues. 

                      node1                                                     node2                 
                                                                                                                          
                      eaa                                                                                              
                                                                                                                         
         producer1      consumer1                        producer2    consumer2         

Because edge apps on different edge node all can access service eaa, the consumer can consume the service provided by producer which is on a different node.

For example:
producer1 is located in node1 and consumer2 is located on node2. The networking flow will be:
   producer1 -> service eaa -> pod eaa
   consumer2 -> service eaa -> pod eaa
                   node1                                                     node2                 
                                                                                                                          
                      eaa                                                                                              
                                                                                                                         
                producer1                                             consumer2     

OS (Ubuntu)

OpenNESS only supports Centos but ICN is based on Ubuntu 18.04. This gap analysis is to investigate how to deploy OpenNESS on Ubuntu 18.04

OpenNESS only supports Centos but ICN is based on Ubuntu 18.04. By changing the ansible scripts of OpenNESS, it is able to deploy OpenNESS on Ubuntu 18.04.  The following parts of ansible scripts need to change:
1.  Following ansible roles can be removed for OpenNESS master: grub, cnca, multus, nfd. Ansible role grub can be removed for OpenNESS node. Because:

• grub is used to add hugepages to grub and hugepages are not useful for integration OpenNESS with ICN.
• cnca is not required for integration.
• multus has already been integrated with ICN.
• nfd will be integrated directly with ICN.

2. Centos uses yum to install packages and we need to use apt for Ubuntu.

3. Some packages which will be installed by ansible scripts should be removed or replaced:

• Some Centos packages doesn't exist on Ubuntu and these packages should be removed. For example, yum-utils, device-mapper-persistent-data.
• Some Centos packages' name are different for Ubuntu. For example, python2-pip should be replaced with python-pip, python-devel should be replaced with python-dev.

4. Selinux is not used on Ubuntu and need to remove the ansible scripts configuring selinux.

5. Epel repository is for Centos and Ubuntu doesn't need this repository.

6. Proxy will be set for yum and need to change the scripts to set proxy for apt.

7. Docker installation for Centos and Ubuntu are different. Need to change the scripts following the installation guide. For example: the docker repository is different for Centos and Ubuntu.

8. Auditd is used for Docker. Auditd is delivered with Centos by default but Ubuntu needs to install auditd.

9. Kubernetes installation for Centos and Ubuntu are different. Need to change the scripts following the installation guide. For example: gpg key is different for Centos and Ubuntu, ubuntu use deb and Centos uses repository.

10. cgroups driver is different for Centos (systemd) and Ubuntu (cgroups). By default, cgroups driver is cgroups and need to remove the ansible scripts which configures cgroups driver to systemd.

11. firewalld is used in Centos and need to change to ufw which is used by Ubuntu.

12. Packages are different for installing openvswitch and ovn. Centos uses RPMs. Ubuntu uses openvswitch-switch, ovn-common, ovn-central and ovn-host.

13. Topology manager and CPU manager is configured for edge node's kubelet. No need to use topology manager and can remove these.