harbor 能力测试

测试环境

测试机型:

harbor机型:I2 IO优化型, 4C8G
测试镜像大小: 1.7G
测试工具:harbor perftest

测试集群:

准备k8s集群,通过 k8s node 节点拉取harbor镜像,

准备测试镜像:

1
2
3
docker pull garima0079/bigimage
docker tag garima0079/bigimage {harbor LB地址}/bigimage
docker push {harbor LB地址}/bigimage

测试工具准备:

1
2
3
git clone git@github.com:seamounts/perftest.git -d $GOPATH/src
cd $GOPATH/src/perftest
go build

准备测试集群node docker daemon 配置:

批量更新 k8s node docker daemon,将 harbor lb地址加入到 insecure-registries 列表

1
$GOPATH/src/perftest/cmd/harbor/script/prepare.sh

清除镜像缓存:

批量清除 k8s node 节点镜像缓存

1
$GOPATH/src/perftest/cmd/harbor/script/clear-image-cache.sh {harbor LB地址}/bigimage

测试指标

镜像无缓存并发拉取大镜像的平均耗时,99值耗时

harbor 实例数和拉取时间的关系

harbor 支持的最大并发

无缓存拉取大镜像测试

2 HA harbor测试

测试环境:
harbor 2个实例,主机I2 IO优化型,4C8G,内网LB,内网KS3

测试数据:
执行: $GOPATH/src//perftest harbor pullimage –image 10.21.1.4/hsxue/bigimage –kubeconfig /root/.kube/config, 在集群所有节点并发拉取镜像。

50 node 并发拉取测试

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
Summary:
Total: 73.3745 secs.
Slowest: 73.3734 secs.
Fastest: 0.8629 secs.
Average: 49.8971 secs.
Response time histogram:
0.8629 [1] |∎
8.1140 [0] |
15.3650 [0] |
22.6161 [0] |
29.8671 [0] |
37.1182 [0] |
44.3692 [0] |
51.6203 [31] |∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎
58.8713 [17] |∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎
66.1224 [0] |
73.3734 [1] |∎
Latency distribution:
99% in 57.8476 secs.

100 node 并发拉取测试

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
Summary:
Total: 127.8106 secs.
Slowest: 127.8100 secs.
Fastest: 46.4622 secs.
Average: 70.0965 secs.
Response time histogram:
46.4622 [1] |∎∎
54.5969 [18] |∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎
62.7317 [18] |∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎
70.8665 [19] |∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎
79.0013 [17] |∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎
87.1361 [12] |∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎
95.2709 [8] |∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎
103.4056 [2] |∎∎∎∎
111.5404 [3] |∎∎∎∎∎∎
119.6752 [1] |∎∎
127.8100 [1] |∎∎
Latency distribution:
99% in 127.8100 secs.

200 node 并发拉取测试

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
Summary:
Total: 180.1791 secs.
Slowest: 180.1777 secs.
Fastest: 53.9234 secs.
Average: 112.8468 secs.
Response time histogram:
53.9234 [1] |∎
66.5489 [10] |∎∎∎∎∎∎∎∎∎∎∎∎
79.1743 [16] |∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎
91.7997 [28] |∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎
104.4252 [30] |∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎
117.0506 [32] |∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎
129.6760 [23] |∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎
142.3015 [21] |∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎
154.9269 [20] |∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎
167.5523 [15] |∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎
180.1777 [4] |∎∎∎∎∎
Latency distribution:
99% in 173.5462 secs.

300 node 并发拉取测试

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
Summary:
Total: 246.9771 secs.
Slowest: 246.9760 secs.
Fastest: 53.7463 secs.
Average: 152.4312 secs.
Response time histogram:
53.7463 [1] |
73.0693 [10] |∎∎∎∎∎∎∎∎
92.3923 [22] |∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎
111.7152 [37] |∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎
131.0382 [34] |∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎
150.3612 [34] |∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎
169.6841 [50] |∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎
189.0071 [34] |∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎
208.3301 [42] |∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎
227.6530 [21] |∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎
246.9760 [15] |∎∎∎∎∎∎∎∎∎∎∎∎
Latency distribution:
99% in 237.6006 secs.

400 node 并发拉取测试

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
Summary:
Total: 297.7355 secs.
Slowest: 297.7326 secs.
Fastest: 60.6159 secs.
Average: 187.1423 secs.
Response time histogram:
60.6159 [1] |
84.3275 [7] |∎∎∎∎∎
108.0392 [37] |∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎
131.7509 [41] |∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎
155.4626 [40] |∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎
179.1742 [53] |∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎
202.8859 [54] |∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎
226.5976 [53] |∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎
250.3092 [45] |∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎
274.0209 [36] |∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎
297.7326 [31] |∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎
Latency distribution:
99% in 295.2425 secs.

4 HA harbor测试

测试环境:
harbor 4个实例,主机I2 IO优化型,4C8G,内网LB,内网KS3

测试数据:
执行: $GOPATH/src//perftest harbor pullimage –image 10.21.1.4/hsxue/bigimage –kubeconfig /root/.kube/config, 在集群所有节点并发拉取镜像。

50 node 并发拉取测试

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
Summary:
Total: 68.5314 secs.
Slowest: 68.5308 secs.
Fastest: 41.2531 secs.
Average: 45.7705 secs.
Response time histogram:
41.2531 [1] |∎
43.9809 [27] |∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎
46.7086 [11] |∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎
49.4364 [4] |∎∎∎∎∎
52.1642 [2] |∎∎
54.8919 [1] |∎
57.6197 [1] |∎
60.3475 [0] |
63.0752 [0] |
65.8030 [1] |∎
68.5308 [1] |∎
Latency distribution:
99% in 64.3012 secs.

100 node 并发拉取测试

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
Summary:
Total: 115.8375 secs.
Slowest: 115.8369 secs.
Fastest: 46.4803 secs.
Average: 70.5216 secs.
Response time histogram:
46.4803 [1] |∎∎
53.4160 [20] |∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎
60.3516 [12] |∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎
67.2873 [12] |∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎
74.2229 [15] |∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎
81.1586 [11] |∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎
88.0943 [11] |∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎
95.0299 [8] |∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎
101.9656 [7] |∎∎∎∎∎∎∎∎∎∎∎∎∎∎
108.9012 [0] |
115.8369 [3] |∎∎∎∎∎∎
Latency distribution:
99% in 115.8369 secs.

300 node 并发拉取测试

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
Summary:
Total: 240.1224 secs.
Slowest: 240.1149 secs.
Fastest: 53.0441 secs.
Average: 138.6291 secs.
Response time histogram:
53.0441 [1] |
71.7512 [15] |∎∎∎∎∎∎∎∎∎∎
90.4583 [27] |∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎
109.1654 [40] |∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎
127.8724 [55] |∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎
146.5795 [38] |∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎
165.2866 [32] |∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎
183.9937 [41] |∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎
202.7008 [27] |∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎∎
221.4078 [18] |∎∎∎∎∎∎∎∎∎∎∎∎∎
240.1149 [6] |∎∎∎∎
Latency distribution:
99% in 237.0826 secs.

harbor 性能总结

2 HA harbor 性能

2.1 并发拉取镜像数和拉取延迟关系

1

拉取 1.7G 大镜像所需要的时间和并发拉取node的数量基本成线性正比的关系,并发度每增加100,拉取时间就相应增加一分钟。

2.2 并发拉取镜像数和cpu 使用率关系

2

并发拉取量每增加 100,cpu 使用率就增加 10%,当并发达到 500 时,cpu使用率为50%左右 。可知当拉取镜像的并发数达到500以上时,对 harbor 的压力会比较大。

推荐增加harbor 实例个数来缓解压力。

2 HA harbor 和 4HA harbpr 拉取镜像平均时间对比

50 nodes 100 nodes 300 nodes
2HA harbor 49.8971 s 70.0965 s 152.4312 s
4HA harbor 45.7705 s 70.5216 s 138.6291 s

增加 harbor 实例个数,拉取镜像所花费时间没有明显的的减少。时间增加应该和后端存储的性能和网络带宽有关。