[Opensim-dev] Modifying the networking stack

[Diva Canto]

On 11/14/2014 6:23 AM, Michael Heilmann wrote:
> Thanks for the responses.  I'll go into a little more detail:
>
> We have been running several profilers against OpenSimulator on the 
> MOSES grid, and on my development machine.  The tests were to examine 
> the loading on the server under several different loads, specifically 
> mesh and physics loads.  What we found appears to be that no matter 
> what kind of load we placed on the region, even to the point of 
> becoming unresponsive due to physics and mesh, that scripting and 
> physics load were nowhere near the amount of time spent in 
> OpenSim.Region.ClientStack.LindenUDP once we had more than one or two 
> avatars logged in.  We know from previous investigations at our 
> firewall that network traffic for OpenSim is not that heavy, 
> especially with low numbers of users.

If this is a problem, and you are running a recent-ish version of core 
OpenSim, it sounds like some misconfiguration somewhere. Back in the 
summer of 2013 we had a problem with the server running OSCC'13; the 
kernel was configured to run in some sort of special mode that was 
making everything run badly and unpredictably. We fixed the kernel 
configuration, and suddenly things started running much more smoothly-- 
I don't remember the details, but Nebadon may clarify things.

OpenSim these days can handle 50 people on a single simulator without 
much trouble. If you look at figure 7 of my paper 
(http://www.ics.uci.edu/~lopes/documents/summersim14/gabrielova_lopes_preprint.pdf) 
you will see the quantification of "without much trouble." I suggest 
that you reproduce my experimental conditions with pCamBot and check 
whether your numbers are very different from ours. If they are very 
different, then there's definitely something odd in your setup, as we 
were able to reproduce these numbers in several machines. Feel free to 
contact me directly for details about pCamBot configuration.

Bots aren't real viewers, but they are much better for measuring things 
systematically and detecting problems and bottlenecks than relying on 
real users driven by real people. The performance you get with pCamBot 
will be correlated with the performance you get with real users.


> I ran several Wireshark captures against a Firestorm viewer logging 
> into the MOSES public grid ABWIS region, where we hold our office 
> hours.  I saw that with our current configuration, all traffic between 
> the server and my client, with the exception of http CAPS and fsapi 
> calls, were UDP traffic.  This is not immediately concerning, as we 
> have simian serve our mesh and textures directly. The messages are 
> mostly binary information, so I could not examine closely, but I did 
> see a lot of messages containing identical ASCII strings, such as the 
> name of my avatar.

Hard to say what you saw, but I bet those are the AgentUpdate messages 
that I mentioned before. The viewer sends at least 10/sec. At points, 
the viewer sends much more than 10/sec, up to 60/sec. Again, take a look 
at my paper for understanding what those are, and how OpenSim deals with 
them since OSCC'13.

As I said before, it would be nice to understand why the viewer is so 
eager to blabber its status to the server when nothing is going on.


> My primary concern is the amount of time spent handling networking, 
> not necessarily the networking its-self.  But there is at least a 
> portion of messages on the UDP pipeline that are either reliable, or 
> perhaps should be; and re-implementing a reliable transport over udp 
> introduces load at the application layer, instead of letting a 
> low-level reliable transport such as tcp handle it.  I went to 
> university with a guy who implemented a java networking library 
> completely over UDP, believing that it was faster than a normal TCP 
> socket; but he was neglecting that the networking hardware handles the 
> ACK and retransmission transparently, and without needing for the 
> messages to be handled manually by the application.
>
> This may just be my opinion, but since I was going to be ecamining the 
> network stack anyways, and typically in a client-server scenario the 
> ability to maintain a persistent reliable connection where the server 
> can push important events to the client, that it would be a good 
> idea.  The points about network throttling and QoS are taken, but 
> wouldn't they also typically affect the UDP stream? Working on MOSES I 
> have plenty of problems dealing with external users who operate on 
> restricted networks, and they cannot see traffic aside from 80 and 443 
> without dealing with their own IT personnel.  The fact that it is HTTP 
> over TCP instead of raw TCP makes no difference once it is on a 
> non-standard HTTP port.
>
> I agree that it would be more prudent to look at improving the 
> websocket code and the http server, rather than replace it with a raw 
> TCP socket, especially given that there are multiple plugins, such as 
> jsonsimstats, that use the http functionality directly.
>
> I hope that explains my position a little better.  I would love to 
> hear if there are other plans/ideas in the community to address 
> time-sinks like this one, networking simply appears to us as a good 
> starting point to increase performance and scalability of the system.
>