Looking at the performance cost of chaining multiple web servers versus having a single one.
We have 3 scenarios we're looking at:
- nginx -> haproxy -> backend –
docker-compose up nginx-haproxy - haproxy -> backend –
docker-compose up haproxy - nginx -> backend –
docker-compose up nginx
And we want to understand the performance differences (if any).
> wrk2 -t 50 -c 100 -d 60s -R 200 http://localhost:8080nginx -> haproxy -> backend . This is something like the existing setup.
Thread Stats Avg Stdev Max +/- Stdev
Latency 47.65ms 12.69ms 98.37ms 70.26%
Req/Sec 3.88 6.35 22.00 75.35%
12008 requests in 1.00m, 3.03MB read
Requests/sec: 200.02
Transfer/sec: 51.76KB
haproxy -> backend . This is a proposed optimisation to move session ID concerns out of nginx and just use haproxy.
Thread Stats Avg Stdev Max +/- Stdev
Latency 14.21ms 8.74ms 66.94ms 70.99%
Req/Sec 4.11 13.36 100.00 90.51%
12015 requests in 1.00m, 2.98MB read
Requests/sec: 200.17
Transfer/sec: 50.82KB
nginx -> backend . This is the other proposed optimisation to move routing concerns out of haproxy and just use nginx.
Thread Stats Avg Stdev Max +/- Stdev
Thread Stats Avg Stdev Max +/- Stdev
Latency 27.82ms 7.61ms 56.22ms 73.75%
Req/Sec 3.97 9.25 36.00 85.30%
12026 requests in 1.00m, 3.04MB read
Requests/sec: 200.32
Transfer/sec: 51.83KB
You will get different numbers when you run this. I got different numbers each time!
This was an interesting exercise. Using Lua in haproxy and openresty nginx is a different experience. Openresty comes with lots out of the box, whereas haproxy needed more work – see the Dockerfile for how I installed a library to get openssl functionality available in Lua in haproxy.