Add benchmarking script

[MarkusSintonen]

Summary

Adds benchmarking script comparing httpcore against another popular http library aiohttp. Related to performance issues httpcore has encode/httpx#3215.

This runs 500 requests with 20 requests concurrently. The test is repeated 10 times. Test server has simulated latency of 10ms with 2KB response. Simulating a good network connection (eg in data center conditions).

Checklist

• I understand that this PR may be closed in case there was no previous discussion. (This doesn't apply to typos!)
• I've added a test for each change that was introduced, and I tried as much as possible to make a single atomic change.
• I've updated the documentation accordingly.

[tomchristie]

Great thanks. Is there any sense in us also having the same benchmark for the threaded case?
Eg. Comparison between requests and httpx.

[MarkusSintonen]

Great thanks. Is there any sense in us also having the same benchmark for the threaded case? Eg. Comparison between requests and httpx.

It might make sense for checking the pool optimization with the sync case

[MarkusSintonen]

@tomchristie I added now here the sync benchmark using requests library as a baseline. Rerun the benchmark for the connection pool optimization #924 and updated the description there. The results look pretty promising for the sync-case. With that httpcore reaches perf levels of requests and is even faster than that.

[karpetrosyan]

I think the most important part of performance improvement is writing effective benchmarks. So, I have two questions here:

• Should we run these two benchmarks in one process, or would we get more precision if we isolate them?
• Should we use requests or urllib3? I suppose httpcore is more akin to an alternative to urllib3 rather than requests :d

[MarkusSintonen]

• Should we run these two benchmarks in one process, or would we get more precision if we isolate them?

It doesnt actually work very well when running in a single process as then the benchmark also includes the server side. The sync benchmark especially breaks as the async server wont be able to process the requests as the sync IO is blocking everything. So it would need also to use some threaded sync server which starts to get complex to my taste just for the benchmarking.

• Should we use requests or urllib3? I suppose httpcore is more akin to an alternative to urllib3 rather than requests :d

Good point, I think it makes sense to use urllib3 as a baseline. I switched now to urllib3 and updated the benchmark results in the connection pool PR.

[tomchristie]

Okay, yes.