This is a Go response to https://medium.com/@rymcol/benchmarks-for-the-top-server-side-swift-frameworks-vs-node-js-24460cfe0beb
Currently only /json handler is implemented. /blog implementation is left as an exercise for readers :)
Go version used - go tip aka go 1.8.
Hosting & Environment
Test server configuration:
$ uname -a
Linux xxx 4.4.0-45-generic #66-Ubuntu SMP Wed Oct 19 14:12:37 UTC 2016 x86_64 x86_64 x86_64 GNU/Linux
$ cat /proc/cpuinfo | grep name | head -1
model name : Intel(R) Xeon(R) CPU E3-1230 v3 @ 3.30GHz
wrk was run from a separate server with identical configuration.
Both servers are connected over 1Gbps network.
Build Time Results
time go build
real 0m0.466s
user 0m0.544s
sys 0m0.068s
This is 68x faster than the best build time for Swift code - 32 seconds.
Memory Usage
- After process start: 5Mb
- Peak memory usage for 10 concurrent clients: 9.5Mb
- Peak memory usage for 100 concurrent clients: 10Mb
Compare to Swift memory usage.
Thread usage
It has little sense with GOMAXPROCS :)
JSON benchmarks
$ wrk -t 1 -c 20 -d 60s --latency http://10.10.xx.xxx:8090/json
Running 1m test @ http://10.10.xx.xxx:8090/json
1 threads and 20 connections
Thread Stats Avg Stdev Max +/- Stdev
Latency 248.10us 660.77us 29.34ms 98.64%
Req/Sec 95.98k 4.22k 104.16k 85.02%
Latency Distribution
50% 176.00us
75% 222.00us
90% 346.00us
99% 0.97ms
5739608 requests in 1.00m, 1.96GB read
Requests/sec: 95501.72
Transfer/sec: 33.33MB
Results: 95500 qps, 248us average latency. This is 17x faster than the fastest Swift code with 5.7K qps and 3520us average latency :)
20 concurrent connections is laughable, so below are Go results for 20K concurrent connections:
$ wrk -t 1 -c 20000 -d 60s --latency http://10.10.xx.xxx:8090/json
Running 1m test @ http://10.10.xx.xxx:8090/json
1 threads and 20000 connections
Thread Stats Avg Stdev Max +/- Stdev
Latency 142.28ms 54.24ms 1.01s 85.73%
Req/Sec 76.41k 12.74k 109.64k 68.27%
Latency Distribution
50% 127.46ms
75% 136.15ms
90% 224.86ms
99% 358.61ms
4533633 requests in 1.00m, 1.54GB read
Requests/sec: 75427.92
Transfer/sec: 26.32MB
Peak memory usage for 20K concurrent connections: 500Mb.
Just for fun: nodejs performance on the same hardware
Node.js code has been taken from the original benchmark. Nodejs version:
$ nodejs --version
v4.2.6
20 concurrent connections:
$ wrk -t 1 -c 20 -d 60s --latency http://10.10.xx.xxx:8091/json
Running 1m test @ http://10.10.xx.xxx:8091/json
1 threads and 20 connections
Thread Stats Avg Stdev Max +/- Stdev
Latency 2.75ms 378.70us 24.16ms 93.40%
Req/Sec 7.32k 431.18 8.05k 76.21%
Latency Distribution
50% 2.70ms
75% 2.85ms
90% 3.02ms
99% 3.56ms
437404 requests in 1.00m, 170.57MB read
Requests/sec: 7277.96
Transfer/sec: 2.84MB
20K concurrent connections:
$ wrk -t 1 -c 20000 -d 60s --latency http://10.10.xx.xxx:8091/json
Running 1m test @ http://10.10.xx.xxx:8091/json
1 threads and 20000 connections
Thread Stats Avg Stdev Max +/- Stdev
Latency 1.34s 423.74ms 2.00s 65.37%
Req/Sec 9.33k 3.71k 21.79k 82.67%
Latency Distribution
50% 1.38s
75% 1.67s
90% 1.98s
99% 2.00s
390175 requests in 1.00m, 152.15MB read
Socket errors: connect 0, read 7, write 62627, timeout 131594
Requests/sec: 6493.70
Transfer/sec: 2.53MB
Pay attention to latencies and the errors line :)
Peak memory usage for 20K concurrent connections: 700Mb.
Conclusions
Go is much faster than Swift and node.js. And it uses less memory :)