On Sun, 2011-04-17 at 07:35 -0700, Denis Tarasov wrote: > Thanks for the info. BTW, how scalable the Tundra solution is? How big > virtual world it can support and how many concurrent connections are > possible? Are there any test results somewhere?
No work has been done yet to make it especially scalable. Apart from it being c++, using bullet for physics and being the quite performant Naali codebase alltogether. What do you mean with 'big'? There are no size limits, as in space, but that's not very meaningful 'cause also with size limits (like slviewers 256m regions) you can just put a lot of small things there and get a big world :) I figure you mean number of objects or something more meaningful like that. That is AFAIK not tested, but I think static objects only sitting on the server don't basically 'cause any load. So probably you can make a huge scene where can fly around happily. There is a demo that helps to test that, scenes/Infinite/drive.js .. which generates road & trees ahead of you as long as you move with the cam (press 'w' to go forward). Would be easy to make a procedural generation thing like that to create a huge scene to see what bottlenecks we'd start getting then. A lot of moving objects on the server, combined with many connected clients, is where we have the bottleneck now .. as in taking a lot of network bandwidth. The first planned optimization step is to make custom packets for moving objects. To prepare this, Jukka recently made an analysis of the traffic in comparison to commercial implementations like Unreal Tournament 3: http://wiki.realxtend.org/index.php/A_comparative_analysis_of_Tundra_network_usage Currently Tundra just uses the generic Entity-Component attribute synchronization mechanism also for movement, there are no separate movement messages at all but positions are synched like all scene data. Besides crafting custom messages, Erno has been looking into simply using zlib to compress the traffic. For supporting large worlds, there is no mechanisms yet, but some ideas -- like supporting the normal division into rooms/areas that MMOs do, and I've been learning and keeping track of the Sirikata Meru and that Intel scalability work too. ~Toni > On Apr 15, 7:15 am, Toni Alatalo <t...@playsign.net> wrote: > > Let's all go so we'll get good test results about scalability! Don't > > know if Naali/Tundra works well enough on those regions, will be fun to > > try too, but using some other viewer if not. > > > > Whether is best to use that OpenSim tech for scaling for large events > > (potentially thousands of AVs in single place, hosted by a farm of > > servers) with realXtend (Naali) or if we can adopt the tech to Tundra > > server is to be seen later, but informative test results will be great > > anyway to see what that approach really delivers. > > > > ~Toni > > > > On Thu, 2011-04-14 at 20:34 -0700, Lake, Dan wrote: > > > The virtual worlds team at Intel has been investigating network > > > performance in OpenSim. We recently made changes to prioritization and > > > queuing of scene updates, adaptive client throttles and packet resend > > > timers. Our validation so far has used mostly synthetic workloads but the > > > results indicate greatly improved network performance, especially for > > > high latency and low bandwidth clients. > > > > > We are conducting a networking test in ScienceSim next Tuesday and invite > > > OpenSim users from all geographic locations to join us. We will be > > > running our distributed scene graph (dsg) with a high client/avatar count > > > and logging queues, throttles, and bandwidth characteristics of connected > > > clients. > > > > > When: Tuesday, April 19. 9AM PDT (UTC-7). > > > > > Where: Login to 'welcome0419' region on ScienceSim and teleport to a test > > > region from there. > > > > > IRC: island.sciencesim.com/6667 #perftest > > > > > We'll plan on running for an hour but something will probably break > > > sooner. If you come by, please tell us about your experience on IRC or > > > via email. I'm interested in which viewer you are using and how you are > > > connected to the Internet. > > > > > ScienceSim account creation:http://www.sciencesim.com/simian/. > > > Hypergrid probably works but we have not tested it with the dsg regions. > > > I recommend making an account if possible. > > > > > Hope you can make it. Thanks, > > > > > Dan Lake > > > Intel Labs > > > dan.l...@intel.com > > > > > _______________________________________________ > > > Opensim-dev mailing list > > > opensim-...@lists.berlios.de > > >https://lists.berlios.de/mailman/listinfo/opensim-dev > -- http://groups.google.com/group/realxtend http://www.realxtend.org
