!!! see svanoort/python-client-benchmarks for up to date results
- the changes in this forked have been merged into / rewrote in the svanoort version.
- the results here may have an issue related to this bug as svanoort as noticed pycurl/pycurl#309
Micro-benchmarking of two python HTTP client tools : requests and pycurl.
- create (or not) a virtual environement :
mkdir ve; virtualenv ve; . ./ve/bin/activate pip install -r requirements.txt
Two choices : nginx inside a docker image or a flask server.
Warning : the flask server is much slower than the nginx instance, and doesn't allow to test HTTPS connections.
First time :
sudo server/docker-nginx/build.sh
Run the server (the logs will be shown on the terminal) :
sudo server/docker-nginx/run.sh
The server listens on 8000 (HTTP) and 4443 (HTTPS).
./server/flask/server.py
The server listens on 8000 (HTTP)
./benchmark.pyon another terminal
to see some options :
./benchmark.py --help
or edit the file directly
So to sum up for 10000 requests (time express in second) and response of 20480 bytes :
| HTTP (Flask), total time | HTTP (NGINX), total time | HTTPS (NGINX), total time | HTTPS (NGINX), cpu time | |
|---|---|---|---|---|
| pycurl, saving in a new buffer, connection reuse | 15.7 | 4.1 | 12.8 | 4.9 |
| requests, connection reuse | 38.2 | 23.18 | 30.3 | 27.0 |
| requests, no connection reuse | 44.3 | 39.7 | 122.8 | 63.4 |
| pycurl, saving in a new buffer, no connection reuse | 66.7 | 56.4 | 130.8 | 32.3 |
| pycurl, saving in the same buffer, connection reuse | 21.9 | 4.5 | 13.7 | 4.8 |
| pycurl, no saving, connection reuse | 18.0 | 4.0 | 13.0 | 4.5 |
Note : The "no connection reuse" speed are limited by my laptop CPU (i3-2367M, 1.4Ghz)