Performance Comparison: Openspace vs. Regular sims

One major reason for the huge discrepancy between what the Company says, and what we hear anecdotally now, has to do with the server differences.

Put it this way, when the original warnings were written, it was in regard to class 4 hardware. And on class 4 hardware, it's still quite true.

It's a shame Wildefire didn't test on a class 4 openspace region floating standalone. Unfortunately I doubt there are any; as all class 4's all had to be attached to other regions back then.

On one of those Talarus, your worst nightmares about regions can come true, and then some.

Class 5 openspaces are a different animal. If a class 5 standard region is a lion, an openspace class 5 might be a healthy bobcat. But a class 4 openspace isn't merely a housecat by further comparison - it's more like a hummingbird.

And that's the class of region the warnings were written about, and I have yet to see anything but anecdotal updates. Why change the text about openspaces at all though - it's a great disclaimer!

So what's the great thing about a class 4 openspace? Why did anyone bother to keep them? Well, put it this way, how about a grandfathered 65,000 meters with 3750 prims for fifty bucks a month. About 1/6 the current pricing of a standard private region!

I'd offer one for testing, but I don't have any.

Ah, and incidentally - generally the mainland regions are so choked with prims and textures, temp rezzers and vendor scripts and nonsense.... that a fairly sparse class 5 openspace seems like heaven with the same 20 avatars.

I can't prove it, but I think it has to do with openspaces sharing the same server being totally empty most of the time. Remember how socially dead these regions can be most of the day.

Actually, Desmond, it's a load of crap, for two reasons:

1) Quote Jack Linden: "For owners of private islands that are on class 4 servers, I am afraid that we will not be offering an upgrade path at this time. The performance of a class 5 server, for most uses, is not significantly better than a class 4 machine." -- Emphasis mine.

2) The issues I and other technical-minded folks have experienced has been on NEW supposedly class-5 hardware, and recently.



I only wish that were true, Desmond. However, EXPERIENCE, as well as the Lindens, tell me otherwise.



Because, for the most part, it is true, and was echoed again today at Robin's office hours, by no less than Robin herself!



I am not sure, but the Isle of Wyrms may have a set. The original OS sim set we have *MAY* still be class 4; I would have to check with Daryth on that. She might have decided to upgrade them just to be current with the latest and greatest. Still, you get what you pay for, probably a bit less in some ways, more in others.



*shrug* I know that in my home sim of Great Pubnico, I never EVER have any lag, even when the club across the street has a good 20-25 avatars going full tilt. The only texture spam is from the Adfarmers, and vendors work awesome. Howeer, I admit the region is pretty sparse.



Yeah, but with people buying them up to build clubs, theme parks, event venues, what-have-you, that is going to become more of the norm than what they were originally intended for. People go "ZOMG! CHEAP SIMS!" and go right out and rent one without truly understanding the implications of what an OPEN SPACE sim is supposed to be.

Bah, They had brought my islands over and never told me. I should have checked I guess. Now if I can get the bot guy to tweak the bot to work on the Beta grid...

Holy cow, that was my home sim. Got my First Land (tm) there 3 days after I joined. Was nice but sure went downhill after the casinos moved in. It was what prompted me to buy my first Island. I go back sometimes, but it always makes me sad to see it in the stat it's in.

Hi Wildfire

Just wanted to thank you for taking time out ( I wish i could do the same at times hehe ) to put together this test report .

I'm sure this will help no end of people with questions they have about how Openspaces perform against regular sim's .

/me Stand's up and clap's for Wildfires effort's

Should come back to see it sometime.

Just my mall, a couple clubs, a SWRPG Sith Treasury, a store, and about half a sim of open space punctuated by (mostly absorbed now, thanks to my efforts) adfarms.

Okay I don't understand the second paragraph... I mean I understand it, but it doesn't make sense to me. It's okay to buy a 4k sm parcel on a regular sim and build a house on it or set up a store or club on it.... but on the prim/script performance equivalent of a 16k sm parcel, one *shouldn't* do those things?

Sure you don't want to set up the density that you'd set up on a regular sim, but doesn't the prim limitation pretty much prevent that anyway?

Tell me: is the use of an Openspace Sim different than the use of a 16,000sm parcel on a regular sim?

These results show that it very definitely *is* different. One quarter of a full sim can use the full sim's resources when nobody else is around; it appears from these results that an OpenSpace sim gets about 1/4 of the CPU capacity no matter what's happening on the other sims that share that CPU.

It's more complicated than that, of course, so overall OpenSpace performance isn't just four times as bad as a Full sim. When the load average is very low, performance can even be better (Wildefire suggests perhaps due to a tweaking of script priority or timeslice in the frame), but when load is high, it seems to perform at about 1/3 a Full (perhaps because some of the performance is dependent on non-CPU resources that are shared instead of split), and with higher variability (perhaps because of demand for those shared resources by CPU-coresident OpenSpaces).

So to me the simplest explanation of these results is that OpenSpaces use virtualization or the simplest possible quota system for timesharing. I'd hoped that, rather, the entire CPU would be available if there was no competition from the CPU-sharing OpenSpaces, but that you were guaranteed a 25% timesharing quota when there was competition for capacity. That would have made OpenSpaces very similar to using 1/4 of a Full sim--except that you couldn't look on the map to see who else was lagging your performance.

Now, that's just the simplest explanation. A more complicated hypothesis is that OpenSpaces use the more elaborate timesharing of the CPU, but other resources (like memory, perhaps) are allotted just the fixed 25% of total capacity, and the tests just demanded more than the allotment (perhaps triggering paging).

Of course, this is all completely knowable; if one asked the right Linden, they could probably just give us the URL for whichever "virtualization" package they're using.

But after all that thumb-sucking, the real reason I wanted to post was to thank Wildefire (and her contributors) for a hard job well done. The results may have been what others thought they knew, but now they have data. And in the detailed results, there were many things that I, for one, certainly did not know. So, *thanks*!

Whatever about the answers posted in response to this, we now have a great put-down line for someone who is a 'bit too much'.

"You're just a bounding box with some rendering client-side!"

A 100% totally idle, totally empty sim, openspace or not, will still try to run at 45FPS. I know there's at least one Linden looking into changing this so that sim will sorta throttle down if there's nobody on it and there aren't any scripts doing I/O on the sim.

On full sims, sharing a server with an empty sim that's running at 45FPS doesn't really have much impact. On an openspace one, sharing a CPU core with 3 empty openspace sims, each trying to run at 45FPS, is more noticable..

1) Jack Linden is wrong on this one. Great guy, but may simply not with the kind of deal-with-it-every-day experience with these things that a lot of us have.

If anyone says that class 4 and class 5 voids are not that different, then offer to do a region class switch - their class 5 void for your old class 4.

They will be suffering, and quick. You can see the performance differences with Mysti's region performance tool.

Ever wonder why Mystical Cookie originally made that thing? I was bugging her to check into all this, and one of the first things we tested was class 4 and class 5 voids. Ask her yourself.

The following blog post doesn't mention the performance tool testing, but it's pretty clear I had a few people checking this stuff out (understatement) - I ran my tests in Adam Zaius's class 4 voids (Rigel I think the name was, might be wrong on the name it's been almost two years) and Caledon Primverness, a class 5 void.

http://cavorite.hexaflexagon.com/blog/?p=56

Davan's blog also debunks a lot of nonsense about "listener lag" and other technical myths out there. ( http://cavorite.hexaflexagon.com/blog/?p=65 )



2) Utterly irrelevant. Whatever you think of class 5 void performance, it only goes way down on a class 4 void.


All of this is simple enough to check, today. Someone can just go re-run the tests.

* * * * *

Also, that statement from Jack looks old. Last I looked there sure was an upgrade path from class 4 to class 5 standard regions. A charge, a tier increase and a standard support ticket type to ask for it, too.

That's an interesting observation. One thing I've been doing to 'cause lag' is open a lot of listeners (88,000 or so). I found it really didn't cause any.

It really would be more accurate to say the lag comes from the scripts processing of the listener event, rather than the listener itself. Some people have a lot of processing going on in your typical listener event. Having the listener open itself is not laggy. If your Listener event quickly discounts any input as not being meant for that particular object, the additional lag will be minimal.

Haha, I wonder what would happen if you had, say, a hundred listeners running that immediately called llSleep(1000) in the listen event. That would be like a time bomb.

Heh, I might try that actually...

Most of LL's servers have been dual-core machines, Opterons or Xeons. Quad-core servers began shipping in quantity a year ago ... I have a stupid question. Are openspace sims simply quadcore servers with 4 instances of a Second Life server running, each with affinity set to a different core? In other words, 4 sims/CPU, 1 sim per core?

If LL is running a fully virtualized data center, this would of course not be the case ... but Cory always gave the impression that their computing network architecture was quite simple (although presumably within the asset clusters storage at least would be fully virtualized).
.

Class 5's are quadcore servers. 4 normal sims per server or 16 openspace sims. Each core will be running 1 normal sim or 4 openspaces. The 4 cores of any specific server are all supposed to be normal sims or openspaces.

Ah! ty.

And I gather they run sims simply as tasks on their allocated cores, they do not use SMP and let the servers manage the resources? If so, that seems sensible. It would be amazing if in the rush to develop SL they had taken time to structure it well for multiprocessing.
.

So to summarize for someone who is having a bit of difficulty with the technical data here... can the following be said to be true?

If you're talking about script and avatar performance, it's better to have 16,000 sm of land on a full sim, than to have an open space sim... e.g. don't move a club to an openspace sim, or one would get a big performance hit?

That depends entirely on what else may or may not be on the full sim with you.

Well I just meant in "general".... I've had a parcel (less than 16k though) on a full sim for over a year ... and what else is on the sim is constantly changing. Lately the rest of the sim seems to be fairly empty most of the time in terms of avatars. I don't know about scripts though. (How does one check that?)

I've been contemplating purchasing more land on the same sim, but the opportunity to purchase an openspace sim has come up as well. Prim-wise, an openspace makes sense, but if it's going to be laggier, then it doesn't.

The big problem is that you can't say anything "in general". If you own 1/4 of a full sim, but the other 3/4 are not being heavily used, you should stay if you need to run a lot of scripts, but move if you really want the extra space. If anyone in the sim is running a lot of processor intensive scripts, they can interfere with anything running on your 1/4 sim, and it may make more sense to move to an openspace. It really depends completely on how what you've got compares to what you want and why you want it.

Yes, llSleep is rather evil. But in this instance it wouldn't be nearly as bad as say, calculating PI to 8 digits. Having 100 listeners do that every time someone spoke on the Public Channel would probably be more laggy

I asked about this at Andrew's office hour tonight and he sounded a bit surprised. Don't want to quote him here - not sure that's legal - but you can look for the transcript at http://wiki.secondlife.com/wiki/Andrew_Linden before too long. It was towards the end of the meeting - maybe 15:50 or so.

He also confirmed that sim hosts (server hardware) always run either all openspace sims or all full sims - they don't mix sim types on a single server.

Interesting read ..

I've been trying to think of other reasons why the empty openspace test would be consistently faster than the empty regular-sim test after several restarts (server changes) on each sim. Settings tweaks were the first thing that came to mind. Faster hardware for openspace servers is a possibility, but that would require that during my tests the openspace sims were always running on faster servers than the regular sims in all 12 rounds, since sim times were pretty consistent for each configuration, except for the one openspace test that ran 17 minutes faster than the others.

He did say that it was possible for an openspace to get an entire host server to itself while full sims might be sharing the host with other full sims. I don't know enough about how sims get assigned to hosts to say if it's _likely_ that that'll happen, though.

The discussion wasn't very long - maybe 15-20 chat lines total but he very clearly did say that they hadn't done any weird scheduling stuff on openspace sims to make them run scripts faster.

All I'm saying that is that he said this isn't something they did on purpose. That doesn't mean you didn't see what you saw - not at all! It could well be that there's some unrealized aspect of the sim being 1/4-sized that just makes them faster than a full sim, when both are empty.

He's got two office hours a week and the format is pretty open, unless the LL people there have got something they want to talk about - stop by and chat with him!