[Opensim-dev] some scalability tests...

Wed Jan 27 07:58:02 UTC 2010

final entry on this...

we seem to have found a combination of settings, execution environment, and
changes to opensim that give us a reasonably stable experience (there is
nothing "normal" about feeding the viewer 100 regions to render so keep it
all in perspective, its most likely going to break if you hang out there
long enough :-)...

* we got rid of the megaregions. we wanted a "many" region view that we get
with megaregons, but we didn't need the multi-region physics. we also found
that megaregions next to megaregions have some very painful effects on
border crossings especially going west & south. and big megaregions are
worse (as far as i could tell, the viewer thinks it doesn't have connections
outside its 10x10 range and the sim thinks the viewer already has the
connection and then weird stuff happens when the region actually comes in
view). instead of the megaregions we hacked opensim to give out all of the
regions within 5 coordinates of the current region (which seems to be the
max the viewer can handle). that works out to about 100 regions that are
visible. the change we made for the test isn't particularly useful in
general, but i think we can try to get the viewers far clip distance to set
the number of regions to send back.

* 64 bit was causing all kinds of problems with the jpeg decoder (no
surprise). given that the only jpeg being decoded was the map image, there
might be something in the way the map is created that causes the problems.
Regardless... given that every time you opened the map view in the viewer,
the region would grab you and not let go (no further TPs, no border
crossings, etc) it seemed like we needed a different solution. so now we're
running 16 simulators, each 8x8 for a 64 regions per simulator and we're
using the 32 bit launcher to force it to stay in 32 bit mode.

* another thing that caused problems at times was the number of mysql
connections. increasing the number of simulators increases the number of
connections. db contention was causing some bad problems on startup... we
haven't dug into the root cause. instead, we've changed to use an
alternative mysql connection management solution that uses mysql connection
pooling. that seemed to get us past the initial problems (yeah... you can
always up the mysql connections... but more efficient use of the connections
seems like a good idea). john's working on a patch against the current
opensim master for that code.

at any rate... it works. and for proof here's a picture of yellowstone
national park from the top of one of the mountains on the eastern side of
the park:
http://www.sciencesim.com/wiki/lib/exe/fetch.php/yellowstone.png

(or come see it on the scisim grid.., search for the "geography" regions on
the map)

--mic

On Fri, Jan 22, 2010 at 6:01 PM, <diva at metaverseink.com> wrote:

> +100! Really great that Intel is pushing the performance limits...
>
> Diva
> (coming home from an event in the university that had lots of local
> industry, including gaming industry, even more convinced that the key
> piece that is missing in all of it, for the longest time, is... an open
> source MMO platform that *everyone* can explore with)
>
> Kyle wrote:
> > Fantastic work Intel!
> >
> >
> >
> > *From:* opensim-dev-bounces at lists.berlios.de
> > [mailto:opensim-dev-bounces at lists.berlios.de] *On Behalf Of *Mic Bowman
> > *Sent:* Friday, January 22, 2010 7:19 PM
> > *To:* opensim-dev at lists.berlios.de; scisim-discuss at googlegroups.com
> > *Subject:* Re: [Opensim-dev] some scalability tests...
> >
> >
> >
> > Grid mode. Connected to SciSim. Thanks to some help from Brian, we put
> > the Yellowstone National Park terrain on it today. I think Brian said it
> > works out to about a 1:10 scale. With the far view distance its pretty
> > impressive (terrain textures are borked). The performance is terrible,
> > but good enough to look around some. Even a completely empty region
> > consumes resources and 1000 of them consume a LOT of resources. If you
> > come over to sciencesim, look for "geography11 00". I'll probably be
> > restarting the regions to get the map tiles updated & i'm not making any
> > promises to keep it up very long. But it does make for a very neat
> > demonstration and its been a very useful experiment.
> >
> > --mic
> >
> > On Fri, Jan 22, 2010 at 3:27 AM, Impalah Shenzhou <impalah at gmail.com
> > <mailto:impalah at gmail.com>> wrote:
> >
> > Just a stupid silly question: standalone or grid mode? If grid: where
> > were the ugaim servers? Same machine?
> >
> > And a comment: 1024 regions? 8 hours for booting? Weird!
> >
> > 2010/1/22 Mic Bowman <cmickeyb at gmail.com <mailto:cmickeyb at gmail.com>>
> >
> > this is just fyi... and a very positive comment about how far opensim
> > has come in recent months!
> >
> > as part of sizing the hw requirements for a mirror world project we're
> > exploring... we wanted to do some scalability tests on the capacity of
> > individual simulators in terms of the total number of regions. we wanted
> > baseline numbers that focus on just the most simple region
> > configuration: a completely empty region with default terrain. that
> > is... this is JUST simulator overhead.
> >
> > all tests are done on one of our blades... dual proc, quad core with 16G
> > ram running 64bit ubuntu. mono 2.6. and the tests are hosting 1024
> > regions in a 32x32 grid. the simulator configuration was our standard
> > sciencesim config (XEngine, ODE, groups, wind, sun, etc).
> >
> > the first configuration ran all 1024 regions in *one* simulator. i
> > honestly figured this would crash quickly. it didn't. we managed to get
> > all 1024 regions created and running well enough to walk around. it did
> > take almost 8 hours to start. the first couple regions were created in
> > 1-2 seconds each. by the end, it was taking 4-5 minutes per region.
> > clearly there is something quadratic in there (stop using linear lists,
> > they are evil! :-). but it could have been the mono garbage collector.
> > who knows... not sure its worth too much investigation because i can't
> > imagine ever running a config like this for real. the simulator did
> > crash when we opened the map in the viewer. the crash was caused because
> > we ran out of sockets. while it was running, the simulator used just
> > over 10G of ram and was running at about 700% CPU utilization (kind of
> > scary to see load averages in the 900 range!).
> >
> > the second configuration ran 16 simulators each with 64 regions. startup
> > took about 30 minutes (each of the 16 simulators avoided the quadratic
> > "knee" we hit with the one big simulator). consumed about 11G of memory
> > and was again consuming essentially all cycles on the machine
> > (completely idle regions aren't very idle). all 16 simulators died just
> > after startup with a "too many open files" error. not sure what caused
> > it, but all of them died loading the same terrain dll as part of some
> > wildcard function looking for dlls. that one is an interesting bug.
> >
> > the final configuration and the one we're shooting for in the short term
> > runs 16 simulators, each with an 8x8 megaregion. again startup was
> > around 40 minutes. we might be able to cut that time down by starting
> > all 16 simulators simultaneously rather than 4 at a time. i really just
> > wanted to make sure that some of the spikes we see in startup didn't
> > cause failures (some race condition causes all threads to consume
> > maximum processor cycles randomly on startup right now). and, well... it
> > just worked. i figured the viewer would die horribly (it can't handle
> > 250K "real" prims very well) but it survived just fine. turn off far
> > clip with 8x8 megaregions providing neighbors and you are "capable" of
> > contacting simulators in a 24x24 region range!. the view is pretty cool
> > though it seems to not go beyond 12x12. :-) there are a bunch of
> > problems (like you can't see the terrain in the region you're standing
> > in)... but there are a LOT of little humps of terrain in view. oh... and
> > it takes a lot of patience to get everything loaded.
> >
> > so... the conclusion... this opensim thing is pretty amazing! good work
> > everyone.
> >
> > --mic
> >
> > _______________________________________________
> > Opensim-dev mailing list
> > Opensim-dev at lists.berlios.de <mailto:Opensim-dev at lists.berlios.de>
> > https://lists.berlios.de/mailman/listinfo/opensim-dev
> >
> >
> >
> > _______________________________________________
> > Opensim-dev mailing list
> > Opensim-dev at lists.berlios.de <mailto:Opensim-dev at lists.berlios.de>
> > https://lists.berlios.de/mailman/listinfo/opensim-dev
> >
> >
> >
> > No virus found in this incoming message.
> > Checked by AVG - www.avg.com
> > Version: 9.0.730 / Virus Database: 271.1.1/2638 - Release Date: 01/22/10
> > 14:33:00
> >
> >
> > ------------------------------------------------------------------------
> >
> > _______________________________________________
> > Opensim-dev mailing list
> > Opensim-dev at lists.berlios.de
> > https://lists.berlios.de/mailman/listinfo/opensim-dev
> _______________________________________________
> Opensim-dev mailing list
> Opensim-dev at lists.berlios.de
> https://lists.berlios.de/mailman/listinfo/opensim-dev
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://opensimulator.org/pipermail/opensim-dev/attachments/20100126/4706f858/attachment-0001.html>