I was searching for a more efficient server solution as an alternative to Fluentd HTTP endpoint. Since I had previous experience with Go language it seemed to be a good candidate for the job.
The server was supposed to accept JSON events in max. 10-records batch.
Estimated load was around 1500 req/sec.
I used wrk tool for benchmark. It seemed more efficient than Apache Benchmark.
Tests were performed using two linux instances hosted on DigitalOcean.
- 2 Core Processor
- 2GB RAM
- 40GB SSD
- Ubuntu 14.04
After a few tests it turned out, that each instance is able to generate 800 req/sec https traffic at maximum. This can be a bottleneck when testing more efficient backends.
Another very important factor is network. Testing from my local machine I got numbers even 6 times less just because of poor coworking network.
wrk -t12 -c50 -d30s -s post.lua https://example.org
Run on each test instance.
curl -X POST -d post.json -H "Content-Type: application/json" -v https://example.org
Application was hosten on Google App Engine and Google Compute Engine using B2 and n1-highcpu-2 machine types.
- Accept
JSONrequest - Perform
Content-Typeand HTTP Method validation - Parse
JSONcontent - Encode
Metadatapart back toJSON
For HTTP handling I used fasthttp, Gin, Iris, net/http. JSON was handled using encoding/json and ffjson libraries.
Payload was generated using online JSON Generator.
[
{
"id": "e542e0ec-3b3e-43e3-86ae-c48f57b3fe40",
"timestamp": 1460616245,
"metadata": {
"index": 0,
"guid": "42b3cde7-c8b5-4739-a277-7b3b0189054f",
"isActive": true,
"balance": "$2,784.60",
"picture": "http://placehold.it/32x32",
"age": 40,
"eyeColor": "green",
"name": "Daniels Parrish",
"gender": "male",
"company": "KOZGENE",
"email": "danielsparrish@kozgene.com",
"phone": "+1 (968) 562-3911",
"address": "727 Lorimer Street, Newry, Nevada, 2684",
"about": "Tempor proident laboris deserunt esse aute occaecat in sit commodo et mollit sunt. Labore laboris dolore nostrud est voluptate tempor fugiat est commodo voluptate esse. Dolore in sit dolor adipisicing ut officia laborum. Ullamco consectetur laboris amet nostrud deserunt qui aliqua cupidatat officia adipisicing ipsum voluptate. Ut Lorem laborum est velit pariatur ut ut velit ipsum cupidatat.\r\n",
"registered": "2014-01-19T04:20:48 -01:00",
"latitude": -81.924366,
"longitude": 80.839932
}
},
{
"id": "d104c643-9f4f-44a5-9f99-d3603cfc08e7",
"timestamp": 1460616245,
"metadata": {
"index": 1,
"guid": "9a84f2df-6230-422c-982c-b936392be10f",
"isActive": false,
"balance": "$3,216.18",
"picture": "http://placehold.it/32x32",
"age": 21,
"eyeColor": "blue",
"name": "Katie Harrison",
"gender": "female",
"company": "FREAKIN",
"email": "katieharrison@freakin.com",
"phone": "+1 (993) 534-3625",
"address": "198 Jerome Avenue, Thornport, Tennessee, 3677",
"about": "Enim anim officia id nulla. Ipsum commodo qui minim sit veniam exercitation anim esse ut culpa veniam. Sit elit eiusmod voluptate nisi voluptate eiusmod minim consectetur velit dolor est est cupidatat. Reprehenderit anim sit adipisicing veniam sit veniam anim consectetur est. Amet excepteur id proident excepteur id. Consequat aute eiusmod enim minim. Et ea ipsum eiusmod deserunt.\r\n",
"registered": "2016-02-16T07:29:30 -01:00",
"latitude": 3.512001,
"longitude": -140.639882
}
},
...
]
I included Fluentd just for comparison. It wasn't perfectly accurate, because Fluentd had more complicated implementation (e.g. pushed parsed JSON events to BigQuery). I just wanted to see how much Go version is faster and then implement other missing features for stage 2 testing if results were promising.
| Tool | Result |
|---|---|
| fasthttp | 907 req/sec |
| GAE | 407 req/sec |
| Gin | 790 req/sec |
| Iris | 795 req/sec |
| net/http + ffjson | 803 req/sec |
| net/http + encoding/json | 836 req/sec |
| Fluentd | 572 req/sec |
The following results were generated from two tests running simultaneously for 30s.
12 threads and 50 connections
Thread Stats Avg Stdev Max +/- Stdev
Latency 108.42ms 50.11ms 1.35s 94.92%
Req/Sec 37.83 9.85 70.00 72.86%
13522 requests in 30.03s, 3.34MB read
Requests/sec: 450.23
Transfer/sec: 113.88KB
12 threads and 50 connections
Thread Stats Avg Stdev Max +/- Stdev
Latency 104.83ms 28.65ms 328.67ms 81.89%
Req/Sec 38.45 9.64 80.00 73.05%
13742 requests in 30.05s, 3.35MB read
Requests/sec: 457.33
Transfer/sec: 114.33KB
12 threads and 50 connections
Thread Stats Avg Stdev Max +/- Stdev
Latency 124.59ms 53.33ms 702.30ms 91.59%
Req/Sec 33.71 9.91 80.00 77.33%
11835 requests in 30.05s, 2.90MB read
Requests/sec: 393.79
Transfer/sec: 98.83KB
12 threads and 50 connections
Thread Stats Avg Stdev Max +/- Stdev
Latency 123.92ms 53.64ms 704.56ms 91.70%
Req/Sec 34.02 9.66 80.00 79.74%
11904 requests in 30.06s, 2.92MB read
Requests/sec: 396.03
Transfer/sec: 99.39KB
12 threads and 50 connections
Thread Stats Avg Stdev Max +/- Stdev
Latency 120.92ms 34.87ms 564.89ms 83.25%
Req/Sec 33.45 9.64 80.00 77.99%
11928 requests in 30.05s, 2.92MB read
Requests/sec: 396.98
Transfer/sec: 99.63KB
Running 30s test @ https://go-iris-dot-mychat-tracking-beta-1245.appspot.com
12 threads and 50 connections
Thread Stats Avg Stdev Max +/- Stdev
Latency 120.72ms 42.92ms 1.58s 88.30%
Req/Sec 33.57 9.26 80.00 79.17%
11977 requests in 30.04s, 2.94MB read
Requests/sec: 398.65
Transfer/sec: 100.05KB
12 threads and 50 connections
Thread Stats Avg Stdev Max +/- Stdev
Latency 120.51ms 42.41ms 1.30s 88.59%
Req/Sec 33.95 9.46 80.00 79.26%
12020 requests in 30.04s, 2.96MB read
Requests/sec: 400.11
Transfer/sec: 100.81KB
12 threads and 50 connections
Thread Stats Avg Stdev Max +/- Stdev
Latency 119.06ms 34.02ms 351.91ms 84.15%
Req/Sec 34.17 9.52 70.00 78.27%
12111 requests in 30.08s, 2.98MB read
Requests/sec: 402.58
Transfer/sec: 101.43KB
12 threads and 50 connections
Thread Stats Avg Stdev Max +/- Stdev
Latency 126.70ms 132.14ms 1.91s 95.47%
Req/Sec 35.62 10.31 80.00 75.00%
12553 requests in 30.05s, 3.10MB read
Socket errors: connect 0, read 0, write 0, timeout 19
Requests/sec: 417.69
Transfer/sec: 105.65KB
12 threads and 50 connections
Thread Stats Avg Stdev Max +/- Stdev
Latency 128.77ms 141.78ms 1.96s 95.58%
Req/Sec 35.72 10.39 70.00 73.58%
12571 requests in 30.04s, 3.11MB read
Socket errors: connect 0, read 0, write 0, timeout 12
Requests/sec: 418.42
Transfer/sec: 105.83KB
12 threads and 50 connections
Thread Stats Avg Stdev Max +/- Stdev
Latency 91.03ms 99.14ms 1.55s 97.83%
Req/Sec 21.93 12.24 80.00 80.98%
6369 requests in 30.05s, 1.45MB read
Socket errors: connect 0, read 0, write 0, timeout 193
Requests/sec: 211.91
Transfer/sec: 49.25KB
12 threads and 50 connections
Thread Stats Avg Stdev Max +/- Stdev
Latency 86.32ms 67.16ms 1.50s 98.17%
Req/Sec 20.99 11.68 80.00 84.55%
5858 requests in 30.05s, 1.33MB read
Socket errors: connect 0, read 0, write 0, timeout 197
Requests/sec: 194.97
Transfer/sec: 45.32KB
12 threads and 50 connections
Thread Stats Avg Stdev Max +/- Stdev
Latency 168.23ms 48.23ms 1.19s 82.59%
Req/Sec 24.37 9.97 60.00 65.21%
8584 requests in 30.06s, 5.50MB read
Requests/sec: 285.58
Transfer/sec: 187.41KB
12 threads and 50 connections
Thread Stats Avg Stdev Max +/- Stdev
Latency 166.72ms 44.58ms 488.04ms 79.04%
Req/Sec 24.37 9.54 70.00 69.40%
8626 requests in 30.05s, 5.53MB read
Requests/sec: 287.02
Transfer/sec: 188.36KB
I was disappointed by the test results. JSON handling seems to be very slow. It seems to be the weakest Go feature, so often used in REST world. There is an open ticket for this matter.
If you don't need JSON for communication you can consider the following alternatives.
These should be blazing fast according to what I've read.