Hi everyone, I just wanna share results and discuss about some things.
Thanks to Andrei Sochirca and JUV RTMP LoadTester tool I tested red5
deployed on a tomcat6 and want to share the results:
The test suite was:
Small EC2 instance: Ubuntu 9.10 + tomcat 6, using open-jdk.6, accessing
oflaDemo from red5 0.8 war
Running the mentioned tool on another instance (3 agents + 1 coordinator)
The agents get as much as they can from the streaming server, so getting the
watchmen_vp6.flv video, that usually needs from 40Kbyte/s - 50Kbytes/s got
on each connection a speed of 110-130 KBytes/s,
The top connections stablished were 82 (average of 70) as the agents
terminate the connections (when they get the complete file) and try to
establish new ones.
The nice thing is that process only needed like 60 MBytes avg of extra RAM
that when is not streaming. And the procesor % was 30avg (27-33).
So my conclusion is small instances can handle more process but the
bandwidth is not enough.
Small instances can handle 160 concurrent connections on a 512kbps video
file aprox, but it gets laggy
Maybe, just maybe, the BW limit is on the aws LAN, as the 2 instances were
communicating with their private IP's (is this assertion true???)
Next thing, i'm gonna test the same things on a large instance i hope large
instances come with a better bandwidth limit, if the limit is the same, then
I'll try to reserve a public IP and do the same test.
Another weird thing that I found is that when I was evaluating the RTMP vs
RTMPT, on a stream of 18Mbytes aprox, the difference on total traffic for
the RTMP=20.28Mbytes and RTMPT = 25.22. is a difference of 4.9 Mbytes,
thats a lot I think, the tunneling technique overheads a lot the payload.
Also the RTMP gets a lot of malformed rtmp packages, maybe because of the
custom RTMP especification. Gonna try too with FSM and check if some
packages are tagged like that.
José Leonardo Diaz Ordoñez
Ingeniero en Electrónica y Telecomunicaciones
Énfasis en Ingeniería Telemática.
Universidad del Cauca, Popayán -Colombia