Open samsp-msft opened 4 months ago
Where can I see your test code?
BTW, it's good to see that the focus on high cpu usage issues has increased recently. This shows that there are some high-traffic services that are starting to use yarp. I have reported this issue in #2427 and hope to raise this priority.
The tested YARP code in question is here: https://github.com/microsoft/reverse-proxy/tree/main/testassets/BenchmarkApp With the crank configuration here: https://github.com/aspnet/Benchmarks/blob/main/scenarios/proxy.benchmarks.yml
The tested YARP code in question is here: https://github.com/microsoft/reverse-proxy/tree/main/testassets/BenchmarkApp With the crank configuration here: https://github.com/aspnet/Benchmarks/blob/main/scenarios/proxy.benchmarks.yml
Thanks!
I thinking a bit more about the YARP results compared to NGINX, in chatting with @sebastienros, crank is controlling the incomming rate based on a fixed number of connections, and then streams within those connections. We get different results according to the number of connections that are thrown at us. Based on http/1.1 with a 1024 byte payload we get the following results:
You can feed more requests at YARP at once, and get a better mean RPS, at the cost of request latency and working set. We should try some different values for each combination and see what effect that has on the RPS, latency and working set. This should apply to the other proxies as well as they may perform better/worse based on the load.