Skip to content
This repository has been archived by the owner on Jan 5, 2024. It is now read-only.

zevenet/kube-nftlb

Folders and files

NameName
Last commit message
Last commit date

Latest commit

Β 
Β 
Β 
Β 
Β 
Β 
Β 
Β 
Β 
Β 
Β 
Β 
Β 
Β 
Β 
Β 
Β 
Β 
Β 
Β 
Β 
Β 
Β 
Β 
Β 
Β 
Β 
Β 
Β 
Β 
Β 
Β 
Β 

Repository files navigation

kube-nftlb

GoDev reference Go report card License

kube-nftlb is a Kubernetes Daemonset able to communicate the Kubernetes API Server, based on a Debian Buster image with nftlb installed.

It can request information from the API Server such as new, updated or deleted Services/Endpoints, and make rules in nftables accordingly.

Features 🌟

  • βœ… nftables backend, the new packet classification framework that replaces the existing {ip,ip6,arp,eb}_tables infrastructure.
  • βœ… Support for Services and Endpoints.
  • βœ… Annotations can be used to configure Services.
  • βœ… Functional and performance tests.
  • ⏹ (Coming soon) Support for Network Policies.

Prerequisites πŸ“‹

  • Docker
  • Minikube
  • kubectl
  • nftables
  • libnftnl11
  • conntrack

Also, you can run debian_tools_installer.sh as root after a fresh Debian Buster install.

root@debian:kube-nftlb# ./scripts/debian_tools_installer.sh

Installation πŸ”§

# Clone the project
user@debian:~# git clone https://github.com/zevenet/kube-nftlb

# Change directory
user@debian:~# cd kube-nftlb

For development:

# Copy and rename .env.example to .env
user@debian:kube-nftlb# cp .env.example .env

# Generate a random password for nftlb
user@debian:kube-nftlb# NFTLB_KEY=$(base64 -w 32 /dev/urandom | tr -d /+ | head -n 1) ; sed -i "s/^NFTLB_KEY=.*$/NFTLB_KEY=$NFTLB_KEY/" .env

# Change user to root
user@debian:kube-nftlb# su

# Modify scripts permissions to grant root execute access
root@debian:kube-nftlb# chmod +x scripts/*.sh
root@debian:kube-nftlb# chmod +x build.sh

# Build the Docker image with build.sh (prerequisites must be met before this)
root@debian:kube-nftlb# ./build.sh

Deployment πŸš€

  1. Start Minikube without kube-proxy being deployed by default:
root@debian:kube-nftlb# minikube start --vm-driver=none --extra-config=kubeadm.skip-phases=addon/kube-proxy
  1. The cluster needs to apply some settings, and they are inside yaml/. coredns will be able to resolve external hostnames and kube-nftlb will be deployed after running this command:
root@debian:kube-nftlb# kubectl apply -f yaml

If for some reason kube-proxy is running, you have to delete it running the following command:

root@debian:kube-nftlb# ./scripts/remove_kube_proxy.sh

Host settings βš™

We have to remove the chains that kubernetes configures by default. To achieve this we have to stop the kubelet service, add a variable to the configuration file and reactivate the service. Follow the following commands:

# Stop kubelet
root@debian:~# systemctl stop kubelet.service

# Disable iptables rules
root@debian:~# echo "makeIPTablesUtilChains: false" >> /var/lib/kubelet/config.yaml

# Empty tables (don't forget to backup your ruleset)
root@debian:~# nft flush table ip nat
root@debian:~# nft flush table ip filter
root@debian:~# nft delete table ip mangle

# Start kubelet
root@debian:~# systemctl start kubelet.service

If everything has gone well, the kubelet service will not create those tables again. Now you will have to apply some commands to recover the connection with your deployments:

# Add chains to filter table
root@debian:~# nft add chain ip filter POSTROUTING
root@debian:~# nft add chain ip filter INPUT '{ type filter hook input priority filter; policy accept; }'
root@debian:~# nft add chain ip filter FORWARD '{ type filter hook forward priority filter; policy accept; }'
root@debian:~# nft add chain ip filter OUTPUT '{ type filter hook output priority filter; policy accept; }'

# Add chains and rules to nat table
root@debian:~# nft add chain ip nat PREROUTING '{ type nat hook prerouting priority dstnat; policy accept; }'
root@debian:~# nft add chain ip nat POSTROUTING '{ type nat hook postrouting priority srcnat; policy accept; }'
root@debian:~# nft add rule ip nat POSTROUTING oifname != "docker0" ip saddr 172.17.0.0/16 counter masquerade
root@debian:~# nft add chain ip nat INPUT '{ type nat hook input priority 100; policy accept; }'
root@debian:~# nft add chain ip nat OUTPUT '{ type nat hook output priority -100; policy accept; }'

Metrics πŸ“ˆ

kube-nftlb metrics are served in localhost:9195/metrics, although this is subject to change.

Prometheus example

  1. Build a Prometheus Docker image running the next command:
root@debian:kube-nftlb# docker image build prometheus -t prometheus-zevenet
  1. Deploy that container as Daemonset:
root@debian:kube-nftlb# kubectl apply -f prometheus/daemonset.yml

Creating resources ✏

Service

In this section we are going to see the different settings that we can apply to create our service. The first thing we have to know is that it is a service and how we can create a simple one and check that it has been created correctly.

A Service is an abstraction which defines a logical set of Pods and a policy by which to access them. A Service in Kubernetes is a REST object, similar to a Pod. Like all the REST objects, you can POST a Service definition to the API server to create a new instance. The name of a Service object must be a valid DNS label name. For example:

# service.yaml
apiVersion: v1
kind: Service
metadata:
  name: my-service
  labels:
    app: front
spec:
  type: ClusterIP
  selector:
    app: front
  ports:
    - name: http
      protocol: TCP
      port: 8080
      targetPort: 80

This specification creates a new Service object named β€œmy-service”, which targets TCP port 8080 on any Pod with the app=front label.

To apply this configuration and verify that our service has been created we have to use the following commands:

  • Apply the configuration contained within the yaml file.
kubectl apply -f service.yaml
  • Shows all services, a service called "my-service" should appear.
kubectl get services -A
NAMESPACE     NAME         TYPE        CLUSTER-IP      EXTERNAL-IP   PORT(S)                  AGE
default       my-service   ClusterIP   IP_cluster      <none>        8080/TCP                 6m12s

Now we are going to check that after the creation of our Service, our farm has been correctly configured. To do that we need the nftlb env values. You can launch the following command from the kube-nftlb directory:

source .env; curl -H "Key: $NFTLB_KEY" "$NFTLB_PROTOCOL://$NFTLB_HOST:$NFTLB_PORT/farms/my-service--http"
{
        "farms": [
                {
                        "name": "my-service--http",
                        "family": "ipv4",
                        "virtual-addr": "IP",
                        "virtual-ports": "8080",
                        "source-addr": "",
                        "mode": "snat",
                        "protocol": "tcp",
                        "scheduler": "rr",
                        "sched-param": "none",
                        "persistence": "none",
                        "persist-ttl": "60",
                        "helper": "none",
                        "log": "none",
                        "mark": "0x0",
                        "priority": "1",
                        "state": "up",
                        "new-rtlimit": "0",
                        "new-rtlimit-burst": "0",
                        "rst-rtlimit": "0",
                        "rst-rtlimit-burst": "0",
                        "est-connlimit": "0",
                        "tcp-strict": "off",
                        "queue": "-1",
                        "intra-connect": "on",
                        "backends": [],
                        "policies": []
                }
        ]
}

The curl that we have launched returns JSON data with the information configured in our farms.

Deployment

In this section we will see how to create a deployment and how we can assign it to other pods (our service). But first we have to know what a deployment is.

Deployments represent a set of multiple, identical Pods with no unique identities. A Deployment runs multiple replicas of your application and automatically replaces any instances that fail or become unresponsive. In this way, Deployments help ensure that one or more instances of your application are available to serve user requests. Deployments are managed by the Kubernetes Deployment controller.

Deployments use a Pod template, which contains a specification for its Pods. The Pod specification determines how each pod should look like: what applications should run inside its containers, which volumes the Pods should mount, its labels, and more. Let's see an example:

# deployment.yaml
apiVersion: apps/v1
kind: Deployment
metadata:
  name: lower-prio
  labels:
    app: front
spec:
  replicas: 2
  selector:
    matchLabels:
      app: front
  template:
    metadata:
      labels:
        app: front
    spec:
      containers:
      - name: nginx
        image: nginx:alpine

Through the "matchLabels" field we can find the pod of our service. We are going to apply our deployment and check that it has been created correctly.

kubectl -f apply deployment.yaml
kubectl get pods --all-namespaces
NAMESPACE     NAME                             READY   STATUS    RESTARTS   AGE
default       lower-prio-64588d8b49-jjlvm      1/1     Running   0          12s
default       lower-prio-64588d8b49-lvk92      1/1     Running   0          12s

Now we are going to check that after creating our deployment, our farm has the backends configured correctly. We will have as many backends configured as replicas we have specified.

NFTLB_KEY=$(grep 'NFTLB_KEY' .env | sed 's/NFTLB_KEY=//')
curl -H "Key: $NFTLB_KEY" http://localhost:5555/farms/my-service--http
{
        "farms": [
                {
                        "name": "my-service--http",
                        "family": "ipv4",
                        "virtual-addr": "IP",
                        "virtual-ports": "8080",
                        "source-addr": "",
                        "mode": "snat",
                        "protocol": "tcp",
                        "scheduler": "rr",
                        "sched-param": "none",
                        "persistence": "none",
                        "persist-ttl": "60",
                        "helper": "none",
                        "log": "none",
                        "mark": "0x0",
                        "priority": "1",
                        "state": "up",
                        "new-rtlimit": "0",
                        "new-rtlimit-burst": "0",
                        "rst-rtlimit": "0",
                        "rst-rtlimit-burst": "0",
                        "est-connlimit": "0",
                        "tcp-strict": "off",
                        "queue": "-1",
                        "intra-connect": "on",
                        "backends": [
                                {
                                        "name": "lower-prio-64588d8b49-lvk92",
                                        "ip-addr": "IP",
                                        "port": "80",
                                        "weight": "1",
                                        "priority": "1",
                                        "mark": "0x0",
                                        "est-connlimit": "0",
                                        "state": "up"
                                },
                                {
                                        "name": "lower-prio-64588d8b49-jjlvm",
                                        "ip-addr": "IP",
                                        "port": "80",
                                        "weight": "1",
                                        "priority": "1",
                                        "mark": "0x0",
                                        "est-connlimit": "0",
                                        "state": "up"
                                }
                        ]
                }
        ]
}

Setting up annotations for a Service πŸ“Œ

We can configure our service with different settings. In general, to configure our service we will use annotations, a field used in our configuration file yaml. In a few words, annotations are a field that will allow us to enter data outside kubernetes.

Through this field we can configure our service with different values that nftlb supports. For example, we can configure the mode of our service, if our backends have persistence or change our load balancing scheduling. We are going to see all the configuration that we can add using annotations, and then we are going to see a small example of the syntax of our annotations.

How to set up annotations

All we have to do is to add it to the metadata.annotations field in the configuration file of our service. Let's see an example:

# annotations-service.yaml
apiVersion: v1
kind: Service
metadata:
  name: my-service
  labels:
    app: front
  annotations:
    service.kubernetes.io/kube-nftlb-load-balancer-mode: "snat"
    service.kubernetes.io/kube-nftlb-load-balancer-scheduler: "rr"
spec:
  type: ClusterIP
  selector:
    app: front
  ports:
    - name: http
      protocol: TCP
      port: 8080
      targetPort: 80

Mode

We can configure how the load balancer layer 4 core is going to operate. The options are:

  • snat the backend responds to the load balancer in order to send the response to the client.
  • dnat the backend will respond directly to the client, load balancer has to be configured as gateway in the backend server.
  • stlsdnat (Stateless DNAT) the load balancer switch destination address for the backend address and forward it to the backend as DNAT does, but it doesn’t manage any kind of connection information.
  • dsr (Direct Server Return) the client connects to the VIP, then the load balancer changes its destination MAC address for the backend MAC address.
service.kubernetes.io/kube-nftlb-load-balancer-mode: "snat"
service.kubernetes.io/kube-nftlb-load-balancer-mode: "dnat"
service.kubernetes.io/kube-nftlb-load-balancer-mode: "stlsdnat"
service.kubernetes.io/kube-nftlb-load-balancer-mode: "dsr"

Persistence

We can configure the type of persistence that is used on the configured farm. This can be configured in two ways. Via annotations and with the sessionAffinity field.

Through annotations:

  • srcip Source IP, will assign the same backend for every incoming connection depending on the source IP address only.
  • srcport Source Port, will assign the same backend for every incoming connection depending on the source port only.
  • srcmac Source MAC, With this option, the farm will assign the same backend for every incoming connection depending on link-layer MAC address of the packet.
service.kubernetes.io/kube-nftlb-load-balancer-persistence: "srcip"
service.kubernetes.io/kube-nftlb-load-balancer-persistence: "srcport"
service.kubernetes.io/kube-nftlb-load-balancer-persistence: "srcmac"
  • srcip srcport Source IP and Source Port, will assign the same backend for every incoming connection depending on both, source IP and source port.
  • srcip dstport Source IP and Destination Port, will assign the same backend for every incoming connection depending on both, source IP and destination port.
service.kubernetes.io/kube-nftlb-load-balancer-persistence: "srcip srcport"
service.kubernetes.io/kube-nftlb-load-balancer-persistence: "srcip dstport"

# The following annotation must be specified in this case
service.kubernetes.io/kube-nftlb-load-balancer-scheduler: "hash"

Through sessionAffinity field:

  • ClientIP The sessionAffinity ClientIP is equivalent to the "srcip" field in annotations.
  • sessionAffinityConfig Through the "timeoutSeconds" field we can configure the stickiness timeout in seconds.
spec:
  type: ClusterIP
  selector:
    app: front
  sessionAffinity: ClientIP
  sessionAffinityConfig:
    clientIP:
      timeoutSeconds: 10

By default, settings made with annotations have priority, even if the sessionAffinity field is defined. The "stickiness timeout in seconds" cannot be configured via annotations. The default value is chosen unless there is a sessionAffinity field and a sessionAffinityConfig where it will collect the value of that field.

Scheduler

We can configure the type of load balancing scheduling used to dispatch the traffic between the backends. The options are:

  • rr does a sequential select between the backend pool, each backend will receive the same number of requests.
  • symhash balance packets that match the same source IP and port and destination IP and port, so it could hash a connection in both ways (during inbound and outbound).
service.kubernetes.io/kube-nftlb-load-balancer-scheduler: "rr"
service.kubernetes.io/kube-nftlb-load-balancer-scheduler: "symhash"

Helper

We can configure the helper of the layer 4 protocol to be balanced to be used. The options are:

  • none it's the default option.
  • amanda enabling this option, the farm will be listening for incoming UDP packets to the current virtual IP and then will parse AMANDA headers for each packet in order to be correctly distributed to the backends.
  • ftp enabling this option, the farm will be listening for incoming TCP connections to the current virtual IP and port 21 by default, and then will parse FTP headers for each packet in order to be correctly distributed to the backends.
  • h323 enabling this option, the farm will be listening for incoming TCP and UDP packets to the current virtual IP and port.
  • irc enabling this option, the farm will be listening for incoming TCP connections to the current virtual IP and port and then will parse IRC headers for each packet in order to be correctly distributed to the backends.
  • netbios-ns enabling this option, the farm will be listening for incoming UDP packets to the current virtual IP and port and then will parse NETBIOS-NS headers for each packet in order to be correctly distributed to the backends.
  • pptp enabling this option, the farm will be listening for incoming TCP connections to the current virtual IP and port and then will parse the PPTP headers for each packet in order to be correctly distributed to the backends.
  • sane enabling this option, the farm will be listening for incoming TCP connections to the current virtual IP and port and then will parse the SANE headers for each packet in order to be correctly distributed to the backends.
  • sip enabling this option, the farm will be listening for incoming UDP packets to the current virtual IP and port 5060 by default, and then will parse SIP headers for each packet in order to be correctly distributed to the backends.
  • snmp enabling this option, the farm will be listening for incoming UDP packets to the current virtual IP and port and then will parse the SNMP headers for each packet in order to be correctly distributed to the backends.
  • tftp enabling this option, the farm will be listening for incoming UDP packets to the current virtual IP and port 69 by default, and then will parse TFTP headers for each packet in order to be correctly distributed to the backends.
service.kubernetes.io/kube-nftlb-load-balancer-helper: "none"
service.kubernetes.io/kube-nftlb-load-balancer-helper: "amanda"
service.kubernetes.io/kube-nftlb-load-balancer-helper: "ftp"
service.kubernetes.io/kube-nftlb-load-balancer-helper: "h323"
service.kubernetes.io/kube-nftlb-load-balancer-helper: "irc"
service.kubernetes.io/kube-nftlb-load-balancer-helper: "netbios-ns"
service.kubernetes.io/kube-nftlb-load-balancer-helper: "pptp"
service.kubernetes.io/kube-nftlb-load-balancer-helper: "sane"
service.kubernetes.io/kube-nftlb-load-balancer-helper: "sip"
service.kubernetes.io/kube-nftlb-load-balancer-helper: "snmp"
service.kubernetes.io/kube-nftlb-load-balancer-helper: "tftp"

Log

We can define in which netfilter flow in which stage you are going to print logs. The options are:

  • none it's the default option.
  • output for traffic going from the host to the pods.
  • forward for traffic that passes through the host. It can be between two pods or from outside to a pod.
service.kubernetes.io/kube-nftlb-load-balancer-log: "none"
service.kubernetes.io/kube-nftlb-load-balancer-log: "output"
service.kubernetes.io/kube-nftlb-load-balancer-log: "forward"

Benchmarks πŸ“Š

This data can be found at resources/ directory.

Environment

  • Host: Bare metal, not virtualized (single node)
  • CPU: Intel i5 660 (4)
  • OS: Debian GNU/Linux 10.6 (buster) x86_64 (netinst, no DE)
  • Kernel: Linux debian 5.8.0-0.bpo.2-amd64 #1 SMP Debian 5.8.10-1~bpo10+1 (2020-09-26) x86_64 GNU/Linux
  • Memory: 3738MiB
  • How many times was the test repeated: 50
  • Software:
    • Minikube v1.15.0 (flags: --vm-driver=none --extra-config=kubeadm.skip-phases=addon/kube-proxy)
    • Docker v19.03.13
    • iptables v1.8.3
    • nftables v0.9.4
    • kubectl client v1.19.4, server v1.19.4
    • kube-proxy v1.19.4 (iptables mode)
    • kube-nftlb commit 72549dff4c892fcc061033fa508ce0d45a5aa7a9

If we're missing something, open an issue and let us know.

Summary

Tests are based on the number of rules iptables/nftables set after a series of steps. There are scripts that do the heavy lifting for you in tests/performance (make sure to understand the README).

Rule count must be known beforehand. We can measure the time between steps if we know how many rules are set in those steps (example: measure how much time (in ms) does it take to change from X rules to Y rules). This counting could vary between systems, although it's a possibility that we don't expect to happen frequently, it could happen for some reason. It's the reason why tests/performance/expected-rule-count.sh.example exists and why the counting must be tailored to your individual system. This means that you shouldn't expect the exact (read carefully) same results from your testing, as we can't be 100% sure this value doesn't change.

Rules are counted with a single shell command. To find out more about this, see tests/performance/test.sh.

We can't do statistics based a single test, because an unique result isn't meaningful on its own and doesn't account for variation. After repeating the test over and over and storing every result individually, we can calculate statistics from those results (ministat) and draw bar charts and boxplots (gnuplot).

The following sections are extracted from the same data (resources/filtered-results.txt). In conclusion, kube-nftlb (nftables) is several times faster than kube-proxy (iptables) (depends on the case how much).

Comparison of averages

Creating a Service

  • 10 endpoints
user@debian:kube-nftlb# cat resources/filtered-results.txt | awk "/^kube-nftlb {$/,/^}$/" | awk "/^\\treplicas-test-010 {$/{flag=1;next}/^\\t}$/{flag=0}flag" | grep -e "create-service" | sed -e "s/^create-service: //g" -e "s/ ms .*$//g" | ministat -n
x <stdin>
    N           Min           Max        Median           Avg        Stddev
x  50             7            63            33         29.12     16.149354

user@debian:kube-nftlb# cat resources/filtered-results.txt | awk "/^kube-proxy {$/,/^}$/" | awk "/^\\treplicas-test-010 {$/{flag=1;next}/^\\t}$/{flag=0}flag" | grep -e "create-service" | sed -e "s/^create-service: //g" -e "s/ ms .*$//g" | ministat -n
x <stdin>
    N           Min           Max        Median           Avg        Stddev
x  50           121          1084           138         179.4      173.5475

kube-nftlb is 6,16 times faster on average than kube-proxy in this case.

  • 50 endpoints
user@debian:kube-nftlb# cat resources/filtered-results.txt | awk "/^kube-nftlb {$/,/^}$/" | awk "/^\\treplicas-test-050 {$/{flag=1;next}/^\\t}$/{flag=0}flag" | grep -e "create-service" | sed -e "s/^create-service: //g" -e "s/ ms .*$//g" | ministat -n
x <stdin>
    N           Min           Max        Median           Avg        Stddev
x  50             7            18             9          9.96     2.4240294

user@debian:kube-nftlb# cat resources/filtered-results.txt | awk "/^kube-proxy {$/,/^}$/" | awk "/^\\treplicas-test-050 {$/{flag=1;next}/^\\t}$/{flag=0}flag" | grep -e "create-service" | sed -e "s/^create-service: //g" -e "s/ ms .*$//g" | ministat -n
x <stdin>
    N           Min           Max        Median           Avg        Stddev
x  50            83           196           152        152.72     15.130777

kube-nftlb is 15,33 times faster on average than kube-proxy in this case.

  • 100 endpoints
user@debian:kube-nftlb# cat resources/filtered-results.txt | awk "/^kube-nftlb {$/,/^}$/" | awk "/^\\treplicas-test-100 {$/{flag=1;next}/^\\t}$/{flag=0}flag" | grep -e "create-service" | sed -e "s/^create-service: //g" -e "s/ ms .*$//g" | ministat -n
x <stdin>
    N           Min           Max        Median           Avg        Stddev
x  50            11          1073            36         73.88     152.00532

user@debian:kube-nftlb# cat resources/filtered-results.txt | awk "/^kube-proxy {$/,/^}$/" | awk "/^\\treplicas-test-100 {$/{flag=1;next}/^\\t}$/{flag=0}flag" | grep -e "create-service" | sed -e "s/^create-service: //g" -e "s/ ms .*$//g" | ministat -n
x <stdin>
    N           Min           Max        Median           Avg        Stddev
x  50           121          1073           190        235.74     181.18226

kube-nftlb is 3,19 times faster on average than kube-proxy in this case.

Deleting a Service

  • 10 endpoints
user@debian:kube-nftlb# cat resources/filtered-results.txt | awk "/^kube-nftlb {$/,/^}$/" | awk "/^\\treplicas-test-010 {$/{flag=1;next}/^\\t}$/{flag=0}flag" | grep -e "delete-service" | sed -e "s/^delete-service: //g" -e "s/ ms .*$//g" | ministat -n
x <stdin>
    N           Min           Max        Median           Avg        Stddev
x  50             7            12            10          9.82    0.87341694

user@debian:kube-nftlb# cat resources/filtered-results.txt | awk "/^kube-proxy {$/,/^}$/" | awk "/^\\treplicas-test-010 {$/{flag=1;next}/^\\t}$/{flag=0}flag" | grep -e "delete-service" | sed -e "s/^delete-service: //g" -e "s/ ms .*$//g" | ministat -n
x <stdin>
    N           Min           Max        Median           Avg        Stddev
x  50            21           974           941        900.92     183.39484

kube-nftlb is 91,74 times faster on average than kube-proxy in this case.

  • 50 endpoints
user@debian:kube-nftlb# cat resources/filtered-results.txt | awk "/^kube-nftlb {$/,/^}$/" | awk "/^\\treplicas-test-050 {$/{flag=1;next}/^\\t}$/{flag=0}flag" | grep -e "delete-service" | sed -e "s/^delete-service: //g" -e "s/ ms .*$//g" | ministat -n
x <stdin>
    N           Min           Max        Median           Avg        Stddev
x  50             7            17            10          9.76     1.5059067

user@debian:kube-nftlb# cat resources/filtered-results.txt | awk "/^kube-proxy {$/,/^}$/" | awk "/^\\treplicas-test-050 {$/{flag=1;next}/^\\t}$/{flag=0}flag" | grep -e "delete-service" | sed -e "s/^delete-service: //g" -e "s/ ms .*$//g" | ministat -n
x <stdin>
    N           Min           Max        Median           Avg        Stddev
x  50           658           996           967        960.56     46.946413

kube-nftlb is 98,41 times faster on average than kube-proxy in this case.

  • 100 endpoints
user@debian:kube-nftlb# cat resources/filtered-results.txt | awk "/^kube-nftlb {$/,/^}$/" | awk "/^\\treplicas-test-100 {$/{flag=1;next}/^\\t}$/{flag=0}flag" | grep -e "delete-service" | sed -e "s/^delete-service: //g" -e "s/ ms .*$//g" | ministat -n
x <stdin>
    N           Min           Max        Median           Avg        Stddev
x  50             7            23            10         10.38     2.8348415

user@debian:kube-nftlb# cat resources/filtered-results.txt | awk "/^kube-proxy {$/,/^}$/" | awk "/^\\treplicas-test-100 {$/{flag=1;next}/^\\t}$/{flag=0}flag" | grep -e "delete-service" | sed -e "s/^delete-service: //g" -e "s/ ms .*$//g" | ministat -n
x <stdin>
    N           Min           Max        Median           Avg        Stddev
x  50            21          1301          1007       1027.08     165.96924

kube-nftlb is 98,94 times faster on average than kube-proxy in this case.

Charts: time by endpoints number

  • X axis: how many replicas does the deployment have (10, 50 or 100 replicas)
    • The comparison is better shown by joining the two daemonsets in the same value.
  • Y axis: how much time (in milliseconds) does it take to set rules after creating or deleting a Service

Chart time/by-endpoints-number/create-service

Chart time/by-endpoints-number/delete-service

Charts: rules by count type

  • X axis: what is calculated (creating or deleting a Service for N replicas)
    • The comparison is better shown by joining the two daemonsets in the same value.
  • Y axis: how many rules are set depending of the X axis value

Chart rules/by-count-type/replicas-test-010

Chart rules/by-count-type/replicas-test-050

Chart rules/by-count-type/replicas-test-100