coffeeowl
commented
9 years ago OK, I'll talk to myself here. So far the bottleneck seems to be ejabberd itself. Send queue in sleekxmpp is almost empty all the time, yet it takes ~20s to complete a simple ad-hoc command under what I can call a tiny load - 500 bots running the same command every second.
bear
Hello,
I am testing my server component which underperforms, i.e. a burst of 800 simultaneous requests kills it for sure. So, I have setup a simple load test. 100 bots connect to a server (ejabberd) and send requests to the component. The original requests were Ad-Hoc commands, after the first test I changed it to messages for simplicity.
What I find amusing is that the component can't handle 100 bots. Each bot randomly, in interval 1..5 seconds sends a request. When the component replies with just "YES" string, the response time stays constant and < 100ms, but even simple JSON parsing of a small string (around 100 chars) inside message handler results in an overload, in a couple of minutes response time goes up to 10s and continues to increase.
Another interesting thing is that even after I stop my test and try to query the component myself (in pidgin), it still takes the same amount of time for component to reply, i.e. 10s or more. I though that some queue might be still full, but even after waiting for some time, the response time doesn't go down. Only after component restart timing goes down to < 100ms. I have observed the same behaviour with prosody.
Am I missing something?