Smana · February 29, 2016 08:28
diff --git a/Kubernetes network benchmarks b/Kubernetes network benchmarks

 kubernetes network performance tests
 ========


 These are pretty simple benchmarks.
 The main purpose is to give an overall comparison between 3 network plugins that we currently support on the project [kubespray.io](http://kubespray.io):
 
 * **Flannel** v0.5.5
 * **Weave** v1.4.4
 * **Calico** v0.16.1

 Feel free to  propose different test cases, i'll do my best to add them.
 Tests	 duration : 15 minutes

 ##Testing plateform
 ![enter image description here](http://s21.postimg.org/yc6a4pshj/kubernetes_bench_arch.png)


 * the kube-proxy is configured with **iptables** mode which performs better.
 * the loader server uses the **nodeport** to access to a kubernetes service

 ## Throughput
 This is the simpliest test. We just want to get close to the baremetal performance (e.g. saturate the 10g nic).
 the **iperf** client is used from the loader sever with the following options
 ```
 iperf3 -c 10.69.1.45 -p 32105 -t 900 -P 12 -w 32M
 ```

 The iperf server is run within a *pod*.

 2 tests are relevant for the throughput:

 * **direct**: running the iperf server in order to connect to the node where the pod (iperf server) resides.
  e.g. the traffic goes directly to the local pod, no forwarding to another node.
 ![enter image description here](http://s16.postimg.org/ps0c0h1ed/Throughput_direct.png)
 
 * **forward**: running the iperf client against another node. Therefore there's a forwarding traffic to the where the pod is located.
 ![enter image description here](http://s21.postimg.org/5isgsq9s7/Throughput_forward_between_nodes.png)

 ##Requests per second (Nginx)

 The replication controller is composed of 2 replicas.
 Each nginx replicas is configured with 8 [workers processes](http://nginx.org/en/docs/ngx_core_module.html#worker_processes).
 These pods use almost all server's resources where they resides

 The http performance tool used is [wrk](https://github.com/wg/wrk) which has multithreading support.
 below is the command line used for these tests
 ```
 wrk -c 200 -d 900 -t 2 http://10.69.1.45:32222
 ```
 ![enter image description here](http://s27.postimg.org/ssd04p8xv/Nginx_requests_per_seconds.png)


 ##Conclusion




 ####Nodes configuration
 Hardware : Dell m610 servers
 cpu:
 memory: 
 os: 
 kernel: 
 kernel settings: 
 nic: 
 irqbalance
 The kubernetes cluster was deployed with [kubespray.io](http://kubespray.io)

	kubernetes network performance tests
	========


	These are pretty simple benchmarks.
	The main purpose is to give an overall comparison between 3 network plugins that we currently support on the project [kubespray.io](http://kubespray.io):

	* Flannel v0.5.5
	* Weave v1.4.4
	* Calico v0.16.1

	Feel free to propose different test cases, i'll do my best to add them.
	Tests duration : 15 minutes

	##Testing plateform
	![enter image description here](http://s21.postimg.org/yc6a4pshj/kubernetes_bench_arch.png)


	* the kube-proxy is configured with iptables mode which performs better.
	* the loader server uses the nodeport to access to a kubernetes service

	## Throughput
	This is the simpliest test. We just want to get close to the baremetal performance (e.g. saturate the 10g nic).
	the iperf client is used from the loader sever with the following options
	```
	iperf3 -c 10.69.1.45 -p 32105 -t 900 -P 12 -w 32M
	```

	The iperf server is run within a pod.

	2 tests are relevant for the throughput:

	* direct: running the iperf server in order to connect to the node where the pod (iperf server) resides.
	e.g. the traffic goes directly to the local pod, no forwarding to another node.
	![enter image description here](http://s16.postimg.org/ps0c0h1ed/Throughput_direct.png)

	* forward: running the iperf client against another node. Therefore there's a forwarding traffic to the where the pod is located.
	![enter image description here](http://s21.postimg.org/5isgsq9s7/Throughput_forward_between_nodes.png)

	##Requests per second (Nginx)

	The replication controller is composed of 2 replicas.
	Each nginx replicas is configured with 8 [workers processes](http://nginx.org/en/docs/ngx_core_module.html#worker_processes).
	These pods use almost all server's resources where they resides

	The http performance tool used is [wrk](https://github.com/wg/wrk) which has multithreading support.
	below is the command line used for these tests
	```
	wrk -c 200 -d 900 -t 2 http://10.69.1.45:32222
	```
	![enter image description here](http://s27.postimg.org/ssd04p8xv/Nginx_requests_per_seconds.png)


	##Conclusion




	####Nodes configuration
	Hardware : Dell m610 servers
	cpu:
	memory:
	os:
	kernel:
	kernel settings:
	nic:
	irqbalance
	The kubernetes cluster was deployed with [kubespray.io](http://kubespray.io)