Kubernetes网络和集群性能测试

准备

测试环境

在以下几种环境下进行测试:

  • Kubernetes集群node节点上通过Cluster IP方式访问
  • Kubernetes集群内部通过service访问
  • Kubernetes集群外部通过traefik ingress暴露的地址访问

测试地址

Cluster IP: 10.254.149.31

Service Port:8000

Ingress Host:traefik.sample-webapp.io

测试工具

测试说明

通过向sample-webapp发送curl请求获取响应时间,直接curl后的结果为:

  1. $ curl "http://10.254.149.31:8000/"
  2. Welcome to the "Distributed Load Testing Using Kubernetes" sample web app

网络延迟测试

场景一、 Kubernetes集群node节点上通过Cluster IP访问

测试命令

  1. curl -o /dev/null -s -w '%{time_connect} %{time_starttransfer} %{time_total}' "http://10.254.149.31:8000/"

10组测试结果

No time_connect time_starttransfer time_total
1 0.000 0.003 0.003
2 0.000 0.002 0.002
3 0.000 0.002 0.002
4 0.000 0.002 0.002
5 0.000 0.002 0.002
6 0.000 0.002 0.002
7 0.000 0.002 0.002
8 0.000 0.002 0.002
9 0.000 0.002 0.002
10 0.000 0.002 0.002

平均响应时间:2ms

时间指标说明

单位:秒

time_connect:建立到服务器的 TCP 连接所用的时间

time_starttransfer:在发出请求之后,Web 服务器返回数据的第一个字节所用的时间

time_total:完成请求所用的时间

场景二、Kubernetes集群内部通过service访问

测试命令

  1. curl -o /dev/null -s -w '%{time_connect} %{time_starttransfer} %{time_total}' "http://sample-webapp:8000/"

10组测试结果

No time_connect time_starttransfer time_total
1 0.004 0.006 0.006
2 0.004 0.006 0.006
3 0.004 0.006 0.006
4 0.004 0.006 0.006
5 0.004 0.006 0.006
6 0.004 0.006 0.006
7 0.004 0.006 0.006
8 0.004 0.006 0.006
9 0.004 0.006 0.006
10 0.004 0.006 0.006

平均响应时间:6ms

场景三、在公网上通过traefik ingress访问

测试命令

  1. curl -o /dev/null -s -w '%{time_connect} %{time_starttransfer} %{time_total}' "http://traefik.sample-webapp.io" >>result

10组测试结果

No time_connect time_starttransfer time_total
1 0.043 0.085 0.085
2 0.052 0.093 0.093
3 0.043 0.082 0.082
4 0.051 0.093 0.093
5 0.068 0.188 0.188
6 0.049 0.089 0.089
7 0.051 0.113 0.113
8 0.055 0.120 0.120
9 0.065 0.126 0.127
10 0.050 0.111 0.111

平均响应时间:110ms

测试结果

在这三种场景下的响应时间测试结果如下:

  • Kubernetes集群node节点上通过Cluster IP方式访问:2ms
  • Kubernetes集群内部通过service访问:6ms
  • Kubernetes集群外部通过traefik ingress暴露的地址访问:110ms

注意:执行测试的node节点/Pod与serivce所在的pod的距离(是否在同一台主机上),对前两个场景可以能会有一定影响。

网络性能测试

网络使用flannel的vxlan模式。

使用iperf进行测试。

服务端命令:

  1. iperf -s -p 12345 -i 1 -M

客户端命令:

  1. iperf -c ${server-ip} -p 12345 -i 1 -t 10 -w 20K

场景一、主机之间

  1. [ ID] Interval Transfer Bandwidth
  2. [ 3] 0.0- 1.0 sec 598 MBytes 5.02 Gbits/sec
  3. [ 3] 1.0- 2.0 sec 637 MBytes 5.35 Gbits/sec
  4. [ 3] 2.0- 3.0 sec 664 MBytes 5.57 Gbits/sec
  5. [ 3] 3.0- 4.0 sec 657 MBytes 5.51 Gbits/sec
  6. [ 3] 4.0- 5.0 sec 641 MBytes 5.38 Gbits/sec
  7. [ 3] 5.0- 6.0 sec 639 MBytes 5.36 Gbits/sec
  8. [ 3] 6.0- 7.0 sec 628 MBytes 5.26 Gbits/sec
  9. [ 3] 7.0- 8.0 sec 649 MBytes 5.44 Gbits/sec
  10. [ 3] 8.0- 9.0 sec 638 MBytes 5.35 Gbits/sec
  11. [ 3] 9.0-10.0 sec 652 MBytes 5.47 Gbits/sec
  12. [ 3] 0.0-10.0 sec 6.25 GBytes 5.37 Gbits/sec

场景二、不同主机的Pod之间(使用flannel的vxlan模式)

  1. [ ID] Interval Transfer Bandwidth
  2. [ 3] 0.0- 1.0 sec 372 MBytes 3.12 Gbits/sec
  3. [ 3] 1.0- 2.0 sec 345 MBytes 2.89 Gbits/sec
  4. [ 3] 2.0- 3.0 sec 361 MBytes 3.03 Gbits/sec
  5. [ 3] 3.0- 4.0 sec 397 MBytes 3.33 Gbits/sec
  6. [ 3] 4.0- 5.0 sec 405 MBytes 3.40 Gbits/sec
  7. [ 3] 5.0- 6.0 sec 410 MBytes 3.44 Gbits/sec
  8. [ 3] 6.0- 7.0 sec 404 MBytes 3.39 Gbits/sec
  9. [ 3] 7.0- 8.0 sec 408 MBytes 3.42 Gbits/sec
  10. [ 3] 8.0- 9.0 sec 451 MBytes 3.78 Gbits/sec
  11. [ 3] 9.0-10.0 sec 387 MBytes 3.25 Gbits/sec
  12. [ 3] 0.0-10.0 sec 3.85 GBytes 3.30 Gbits/sec

场景三、Node与非同主机的Pod之间(使用flannel的vxlan模式)

  1. [ ID] Interval Transfer Bandwidth
  2. [ 3] 0.0- 1.0 sec 372 MBytes 3.12 Gbits/sec
  3. [ 3] 1.0- 2.0 sec 420 MBytes 3.53 Gbits/sec
  4. [ 3] 2.0- 3.0 sec 434 MBytes 3.64 Gbits/sec
  5. [ 3] 3.0- 4.0 sec 409 MBytes 3.43 Gbits/sec
  6. [ 3] 4.0- 5.0 sec 382 MBytes 3.21 Gbits/sec
  7. [ 3] 5.0- 6.0 sec 408 MBytes 3.42 Gbits/sec
  8. [ 3] 6.0- 7.0 sec 403 MBytes 3.38 Gbits/sec
  9. [ 3] 7.0- 8.0 sec 423 MBytes 3.55 Gbits/sec
  10. [ 3] 8.0- 9.0 sec 376 MBytes 3.15 Gbits/sec
  11. [ 3] 9.0-10.0 sec 451 MBytes 3.78 Gbits/sec
  12. [ 3] 0.0-10.0 sec 3.98 GBytes 3.42 Gbits/sec

场景四、不同主机的Pod之间(使用flannel的host-gw模式)

  1. [ ID] Interval Transfer Bandwidth
  2. [ 5] 0.0- 1.0 sec 530 MBytes 4.45 Gbits/sec
  3. [ 5] 1.0- 2.0 sec 576 MBytes 4.84 Gbits/sec
  4. [ 5] 2.0- 3.0 sec 631 MBytes 5.29 Gbits/sec
  5. [ 5] 3.0- 4.0 sec 580 MBytes 4.87 Gbits/sec
  6. [ 5] 4.0- 5.0 sec 627 MBytes 5.26 Gbits/sec
  7. [ 5] 5.0- 6.0 sec 578 MBytes 4.85 Gbits/sec
  8. [ 5] 6.0- 7.0 sec 584 MBytes 4.90 Gbits/sec
  9. [ 5] 7.0- 8.0 sec 571 MBytes 4.79 Gbits/sec
  10. [ 5] 8.0- 9.0 sec 564 MBytes 4.73 Gbits/sec
  11. [ 5] 9.0-10.0 sec 572 MBytes 4.80 Gbits/sec
  12. [ 5] 0.0-10.0 sec 5.68 GBytes 4.88 Gbits/sec

场景五、Node与非同主机的Pod之间(使用flannel的host-gw模式)

  1. [ ID] Interval Transfer Bandwidth
  2. [ 3] 0.0- 1.0 sec 570 MBytes 4.78 Gbits/sec
  3. [ 3] 1.0- 2.0 sec 552 MBytes 4.63 Gbits/sec
  4. [ 3] 2.0- 3.0 sec 598 MBytes 5.02 Gbits/sec
  5. [ 3] 3.0- 4.0 sec 580 MBytes 4.87 Gbits/sec
  6. [ 3] 4.0- 5.0 sec 590 MBytes 4.95 Gbits/sec
  7. [ 3] 5.0- 6.0 sec 594 MBytes 4.98 Gbits/sec
  8. [ 3] 6.0- 7.0 sec 598 MBytes 5.02 Gbits/sec
  9. [ 3] 7.0- 8.0 sec 606 MBytes 5.08 Gbits/sec
  10. [ 3] 8.0- 9.0 sec 596 MBytes 5.00 Gbits/sec
  11. [ 3] 9.0-10.0 sec 604 MBytes 5.07 Gbits/sec
  12. [ 3] 0.0-10.0 sec 5.75 GBytes 4.94 Gbits/sec

网络性能对比综述

使用Flannel的vxlan模式实现每个pod一个IP的方式,会比宿主机直接互联的网络性能损耗30%~40%,符合网上流传的测试结论。而flannel的host-gw模式比起宿主机互连的网络性能损耗大约是10%。

Vxlan会有一个封包解包的过程,所以会对网络性能造成较大的损耗,而host-gw模式是直接使用路由信息,网络损耗小,关于host-gw的架构请访问Flannel host-gw architecture

Kubernete的性能测试

参考Kubernetes集群性能测试中的步骤,对kubernetes的性能进行测试。

我的集群版本是Kubernetes1.6.0,首先克隆代码,将kubernetes目录复制到$GOPATH/src/k8s.io/下然后执行:

  1. $ ./hack/generate-bindata.sh
  2. /usr/local/src/k8s.io/kubernetes /usr/local/src/k8s.io/kubernetes
  3. Generated bindata file : test/e2e/generated/bindata.go has 13498 test/e2e/generated/bindata.go lines of lovely automated artifacts
  4. No changes in generated bindata file: pkg/generated/bindata.go
  5. /usr/local/src/k8s.io/kubernetes
  6. $ make WHAT="test/e2e/e2e.test"
  7. ...
  8. +++ [0425 17:01:34] Generating bindata:
  9. test/e2e/generated/gobindata_util.go
  10. /usr/local/src/k8s.io/kubernetes /usr/local/src/k8s.io/kubernetes/test/e2e/generated
  11. /usr/local/src/k8s.io/kubernetes/test/e2e/generated
  12. +++ [0425 17:01:34] Building go targets for linux/amd64:
  13. test/e2e/e2e.test
  14. $ make ginkgo
  15. +++ [0425 17:05:57] Building the toolchain targets:
  16. k8s.io/kubernetes/hack/cmd/teststale
  17. k8s.io/kubernetes/vendor/github.com/jteeuwen/go-bindata/go-bindata
  18. +++ [0425 17:05:57] Generating bindata:
  19. test/e2e/generated/gobindata_util.go
  20. /usr/local/src/k8s.io/kubernetes /usr/local/src/k8s.io/kubernetes/test/e2e/generated
  21. /usr/local/src/k8s.io/kubernetes/test/e2e/generated
  22. +++ [0425 17:05:58] Building go targets for linux/amd64:
  23. vendor/github.com/onsi/ginkgo/ginkgo
  24. $ export KUBERNETES_PROVIDER=local
  25. $ export KUBECTL_PATH=/usr/bin/kubectl
  26. $ go run hack/e2e.go -v -test --test_args="--host=http://172.20.0.113:8080 --ginkgo.focus=\[Feature:Performance\]" >>log.txt

测试结果

  1. Apr 25 18:27:31.461: INFO: API calls latencies: {
  2. "apicalls": [
  3. {
  4. "resource": "pods",
  5. "verb": "POST",
  6. "latency": {
  7. "Perc50": 2148000,
  8. "Perc90": 13772000,
  9. "Perc99": 14436000,
  10. "Perc100": 0
  11. }
  12. },
  13. {
  14. "resource": "services",
  15. "verb": "DELETE",
  16. "latency": {
  17. "Perc50": 9843000,
  18. "Perc90": 11226000,
  19. "Perc99": 12391000,
  20. "Perc100": 0
  21. }
  22. },
  23. ...
  24. Apr 25 18:27:31.461: INFO: [Result:Performance] {
  25. "version": "v1",
  26. "dataItems": [
  27. {
  28. "data": {
  29. "Perc50": 2.148,
  30. "Perc90": 13.772,
  31. "Perc99": 14.436
  32. },
  33. "unit": "ms",
  34. "labels": {
  35. "Resource": "pods",
  36. "Verb": "POST"
  37. }
  38. },
  39. ...
  40. 2.857: INFO: Running AfterSuite actions on all node
  41. Apr 26 10:35:32.857: INFO: Running AfterSuite actions on node 1
  42. Ran 2 of 606 Specs in 268.371 seconds
  43. SUCCESS! -- 2 Passed | 0 Failed | 0 Pending | 604 Skipped PASS
  44. Ginkgo ran 1 suite in 4m28.667870101s
  45. Test Suite Passed

从kubemark输出的日志中可以看到API calls latenciesPerformance

日志里显示,创建90个pod用时40秒以内,平均创建每个pod耗时0.44秒。

不同type的资源类型API请求耗时分布

Resource Verb 50% 90% 99%
services DELETE 8.472ms 9.841ms 38.226ms
endpoints PUT 1.641ms 3.161ms 30.715ms
endpoints GET 931µs 10.412ms 27.97ms
nodes PATCH 4.245ms 11.117ms 18.63ms
pods PUT 2.193ms 2.619ms 17.285ms

log.txt日志中还可以看到更多详细请求的测试指标。

kubernetes-dashboard

注意事项

测试过程中需要用到docker镜像存储在GCE中,需要翻墙下载,我没看到哪里配置这个镜像的地址。该镜像副本已上传时速云:

用到的镜像有如下两个:

  • gcr.io/google_containers/pause-amd64:3.0
  • gcr.io/google_containers/serve_hostname:v1.4

时速云镜像地址:

  • index.tenxcloud.com/jimmy/pause-amd64:3.0
  • index.tenxcloud.com/jimmy/serve_hostname:v1.4

将镜像pull到本地后重新打tag。

Locust测试

请求统计

Method Name # requests # failures Median response time Average response time Min response time Max response time Average Content Size Requests/s
POST /login 5070 78 59000 80551 11218 202140 54 1.17
POST /metrics 5114232 85879 63000 82280 29518 331330 94 1178.77
None Total 5119302 85957 63000 82279 11218 331330 94 1179.94

响应时间分布

Name # requests 50% 66% 75% 80% 90% 95% 98% 99% 100%
POST /login 5070 59000 125000 140000 148000 160000 166000 174000 176000 202140
POST /metrics 5114993 63000 127000 142000 149000 160000 166000 172000 176000 331330
None Total 5120063 63000 127000 142000 149000 160000 166000 172000 176000 331330

以上两个表格都是瞬时值。请求失败率在2%左右。

Sample-webapp起了48个pod。

Locust模拟10万用户,每秒增长100个。

locust测试页面

关于Locust的使用请参考Github:https://github.com/rootsongjc/distributed-load-testing-using-kubernetes

参考

基于 Python 的性能测试工具 locust (与 LR 的简单对比)

Locust docs

python用户负载测试工具:locust

Kubernetes集群性能测试

CoreOS是如何将Kubernetes的性能提高10倍的

Kubernetes 1.3 的性能和弹性 —— 2000 节点,60,0000 Pod 的集群

运用Kubernetes进行分布式负载测试

Kubemark User Guide

Flannel host-gw architecture