Why is Second Life so laggy now compared to the past?

[Kia Kiyori]

@Alyona Su But...but...but....4 whole great big monstrous HOURS without SL???

[Scylla Rhiadra]

4 hours ago, ChinRey said:

It's hard to say for sure but this avatar seems to have about 10,000-15,000 triangles, maybe a little bit more but not much. The SL system avatar has  7,186 triangles, the Maitreya/Catwa combo probably somewhere between 150,000 and 200,000 triangles, maybe more and certainly not less. Maitreya and Slink are "best in class" when it comes to performance, most mesh bodies seem to have far more triangles than them.

To get the resolution in KanryDrago's picture, you'd probably need something between 30,000 and 40,000 triangles. If SL had a LoD system for avatars (which it doesn't) we could easily have had avatar triangle counts as high as that with an effective load only marginally higher than the current system avatars.

 

That's mainly about the bones but you do get deformations if the vertice number is too low. We do have such issues with the system avatar but double the vertice/triangle count and there shouldn't be any such problems.

 

No. Well, in theory yes but even the system body has more than enough triangles and vertices to eliminate any such issues.

 

If you're using a Catwa head, it's totally off. Catwa is one of the big "ARC cheaters" and you can safely multiply the figure it shows up with by ten. The Maitreya body on the other hand, is one of the most "honest" fitted mesh items on the market and the actual render load isn't that much higher than what the ARC figure says. Depending on how many layers you use and how much of the body is alpha masked, a Maitreya body may even have an actual render load lower than what the number indicates.

Thank you! These are completely comprehensible answers, and I feel I have a better sense of what is at stake now.

It's really hard to imagine LL trying to introduce a new, sort of intermediary system avatar with, say, 40,000 triangles, in part because I imagine the development time would be prohibitive. But, I suspect, the biggest problem would be adoption. We've grown used to highly detailed models. And while I suspect that it's probably true that, with good quality textures, we might not notice the difference most of the time, there are times when it would be apparent.

It's not just photographers, either: we know that there are people who will actually note on their profiles that they're not going to bother with you unless you have a high-end mesh body. Equally, I was recently alerted to the fact that there are apparently extremely highly detailed girly bits attachments that people like to use for those (*cough*) intimate close-up moments shared with others. (Being something of an ingénue I had no idea. Fascinating.) And indeed, if you look at my pic, you'll notice that, even with a Slink body, there are angular bits around the shoulders. I think that any performance improvements that seem retrograde in terms of graphics quality -- even if the degradation is more imagined than real -- is simply going to be ignored. This is SL, after all: the market place rules.

[Scylla Rhiadra]

2 hours ago, Qie Niangao said:

All those pics look pathetically toony. And at the same time so Uncanny Valley realistic to be squick-worthy. Everybody else's avatar looks like junk, but everybody's own avatar is the pinnacle of verisimilitude. Cuz we've stared at its goofy shape long enough to convince ourselves it's ready for its close-up, Mr. DeMille.

Well, Mr. Niangao, let me tell you that Ces (he said I could call him that!) assured me that I'll be the toast of the Infohubs, and that he's going to find me a good role in his next machinima blockbuster! Maybe even a speaking one!

(And yes, the couch was quite comfortable, but there was a broken spring that kept poking me, somewhere between the 103rd and 104th vertices.)

[Alyona Su]

1 hour ago, AnyaJurelle said:

@Alyona Su But...but...but....4 whole great big monstrous HOURS without SL???

What I did during that time was to open these here forums on me mo-bile fone and annoy all of you!  Hahahaha

Edited August 12, 2019 by Alyona Su

[ChinRey]

1 hour ago, Scylla Rhiadra said:

It's really hard to imagine LL trying to introduce a new, sort of intermediary system avatar with, say, 40,000 triangles, in part because I imagine the development time would be prohibitive. But, I suspect, the biggest problem would be adoption. We've grown used to highly detailed models.

40,000 triangles is actually about the same resolution as our 150,000 triangle mesh avatar/head combo and developing one wouldn't be any different from developing a fitmesh body. How hard it is to implement it is another question. It depends on how much spaghetti there is in the code.

In an ideal avatar focused virtual world I think the solution would be something like:

• A LoD system similar to what we have for rezzed objects but slightly tweaked to fit the size of an avatar as well as possible
• Up to 60,000 triangles for High Lod
• Probably 15,000-25,000 for mid LoD
• 10,000 or less for low LoD
• 5,000 or less for lowest LoD
• An "even lower than lowest" LoD level using standard 12 tri impostors (with server side baked textures) and an adjustable update frequency to replace both the current avatar impostors and the jellydolls.
• More adjustable parameters
• Full access to all 256 possible values of all parameters (some of the old ones coud need a wider range, others smaller steps)
• Recalibration of some of the old parameters
• An "advanced" body modifier option with direct access to the individual "hidden" parameters.
• Re-evaluation of some of the existing parameters, especially the all-but-useless clothes and hair related modifiers.

For something like that, expect one or two months of planning, up to two weeks of modelling and, unfortunately, an unknown but probably very big chunk of time to re-write the software for it.

The end result, if done right, should be:

• Resolution/visual quality at least as good as current fitted mesh body.
• Lag/load maybe a little bit higher than the current system avatar (depending on how much the LoD models could be simplifid without any noticeable effect) but much lower than any mesh avatars.
• Modifiability (is that a word?) covering everything possible with the current system avatar and all existing "standard human" mesh avatars and much more.
• Percise alpha masking for mesh clothes rather than the big chunks mesh bodies have to settle for.
• System clothing that doesn't look like it's painted ontot he avatar. (Not nearly enough flexibility to replace all mesh clothing of course but still enough to cover enough needs to be worth it)
• Very happy mesh clothing creators who only had to relate to a single universal rigging.
• Indifferent specialist mesh body (furries, animals, petites etc., etc.) creators since their products and markets wouldn't be affected very much.
• Very unhappy "standard" mesh body creators since their products and markets would be marginalized over night.
• Very happy, indifferent and very unhappy users for various rational and irrational reasons.

But that's the could-have-been - some would say should-have-been. What will be, is completely different.

Edited August 12, 2019 by ChinRey

[ChinRey]

1 hour ago, Scylla Rhiadra said:

Well, Mr. Niangao, let me tell you that Ces...

Was that Ces Pitt by chance? Just curious.

[KanryDrago]

1 hour ago, ChinRey said:

 

• Very unhappy "standard" mesh body creators since their products and markets would be marginalized over night.
•  

You are assuming that people will switch, personally I doubt it. I have an old body (tmp) because I can't get the look I want precisely on the newer bodies like signature, same reason as some choose other than maitreya bodies. Would I try the linden body yes....do I expect it to be able to do what I want it to? No

[ChinRey]

17 minutes ago, KanryDrago said:

You are assuming that people will switch, personally I doubt it. I have an old body (tmp) because I can't get the look I want precisely on the newer bodies like signature, same reason as some choose other than maitreya bodies.

Yeah but those who choose to stick to their old mesh body weren't going to buy a new one anyway.

We'll never know for sure of course. This was just a thought experiment.

Edited August 12, 2019 by ChinRey

[Theresa Tennyson]

5 hours ago, Qie Niangao said:

On another tangent, I'm still a little curious about the capacity for customization and especially form-preserving animation as triangle counts reduce.

Even though I'm sure my mesh avatar is hauling around way more geometry than he needs, he still deforms ungracefully under animations that sure seem natural enough on the skeleton. But I'm no animator, so maybe I'm misreading the limiting factors.

 

One problem with SL mesh avatars and clothing is that any given vertex can be "influenced" by no more than 4 bones of the skeleton, including the "collision bones" that are used to respond to the "soft tissue" type sliders. The human body and its muscle system is very complex and there are parts of the body that can't be well represented under these conditions - elbows, for instance. 

One of the reasons that I keep using system avatars wearing "standard size" clothing is that since the "standard size" clothing doesn't need to be rigged to collision bones any point can be rigged to more of the main "skeleton" bones and they often seem to move more naturally with some animations.

[KanryDrago]

19 minutes ago, ChinRey said:

Yeah but those who choose to stick to their old mesh body weren't going to buy a new one anyway.

We'll never know for sure of course. This was just a thought experiment.

No I am actually hoping to switch to a newer body, just currently none of them I can adjust to the look I want acceptably. Though I agree most people who want mesh already have a body they like and they won't switch until a body gives them better than what they have. I highly doubt this would be the lab body though. 

[Kyrah Abattoir]

3 hours ago, Scylla Rhiadra said:

Thank you! These are completely comprehensible answers, and I feel I have a better sense of what is at stake now.

It's really hard to imagine LL trying to introduce a new, sort of intermediary system avatar with, say, 40,000 triangles, in part because I imagine the development time would be prohibitive. But, I suspect, the biggest problem would be adoption. We've grown used to highly detailed models. And while I suspect that it's probably true that, with good quality textures, we might not notice the difference most of the time, there are times when it would be apparent.

It's not just photographers, either: we know that there are people who will actually note on their profiles that they're not going to bother with you unless you have a high-end mesh body. Equally, I was recently alerted to the fact that there are apparently extremely highly detailed girly bits attachments that people like to use for those (*cough*) intimate close-up moments shared with others. (Being something of an ingénue I had no idea. Fascinating.) And indeed, if you look at my pic, you'll notice that, even with a Slink body, there are angular bits around the shoulders. I think that any performance improvements that seem retrograde in terms of graphics quality -- even if the degradation is more imagined than real -- is simply going to be ignored. This is SL, after all: the market place rules.

A big problem is that there is a conflation between quality and quantity in SL, and probably a certain amount of misinformation about how meshs are created. A lot of creators in SL are hobbyists, and that's fine, but it also means that they struggle with ideas like:

• How does my work compare to others?
• Have I added enough?

It's not hard to add more polygons, most modeling programs have tools that allow you to programatically smooth models by adding intermediary faces & vertices, you can iterate those steps as much as you want (or as much as your computer is capable of doing), and it will turn a basic 6 sided cube into  a million polygon sphere if you want it to. It's not special and it's not quality, it literally takes seconds.

The workflow for modeling an object is typically a two stage process:

1. Design: Use any tool you want and as much or as little geometry as you need to create the object that you want.
2. Retopologize: Now that you have your design, you remodel it in an optimized fashion (typically in magnet-mode so the new object clings to the reference), making it one single cohesive object. That's where you also take decisions on how you will unwrap it, what should be kept in and what shouldn't.

Some creators kinda 'conflate' both steps, they don't really make complex models anyway so they hardly need to retopologize anyway.

But a sizeable amount of SL creators will go through step 1 and simply skip step 2 because it's a lot of work, their design is "complete" at this point and they want it in SL by now. That's where they end up dumping lods so it will fall under acceptable LI/complexity.

 

Or they will just tell you to suck it up and that "uncompromised art" isn't "low cost"

 

And now for the required pompous quote:
Perfection is achieved, not when there is nothing more to add, but when there is nothing left to take away.  - Antoine de Saint-Exupery

Edited August 12, 2019 by Kyrah Abattoir

[Selene Gregoire]

9 hours ago, ChinRey said:

If you're using a Catwa head, it's totally off. Catwa is one of the big "ARC cheaters" and you can safely multiply the figure it shows up with by ten. The Maitreya body on the other hand, is one of the most "honest" fitted mesh items on the market and the actual render load isn't that much higher than what the ARC figure says. Depending on how many layers you use and how much of the body is alpha masked, a Maitreya body may even have an actual render load lower than what the number indicates.

Would you happen to have any info like this on the VISTA heads? Specfically the Zoe?  I have one I use with my Lara and I'd like to know if they "cheated" or not. I don't believe they did but I'd like that confirmed if possible.

[Suteruni Susanto]

On 8/9/2019 at 6:37 PM, CoffeeDujour said:

They should perform identically. 

The larger inventory will take longer to fetch, but it shouldn't impact performance. It's not your friends list.

Install the Linden viewer, put both main and one alt in the same stock / test Linden avatar and then make sure that they both have identical settings. Take them to a repeatable test location with a minimum number of variables (an empty water sim like KARA) would be best.

Do the same things with the both accounts and make notes.

If there are still substantial differences and you have completely ruled out local settings, start a thread here.

thank you, will try that

[Suteruni Susanto]

On 8/9/2019 at 6:26 PM, Selene Gregoire said:

The accounts are probably on different servers. You can get LL to move the one with the issues but it's going to take a major effort on your part to get it done. 

thanks, yes think you may be correct about LL

[ChinRey]

43 minutes ago, Selene Gregoire said:

Would you happen to have any info like this on the VISTA heads? Specfically the Zoe?  I have one I use with my Lara and I'd like to know if they "cheated" or not. I don't believe they did but I'd like that confirmed if possible.

The "cheat" I was talking about, is one of the three fitmesh LoD bugs. A fitted mesh is of course automatically scaled to fit the avatar. So the size you upload it at doesn't matter to its looks or actual render cost. You can make it as small or big as you like, it's all the same.

The calculated render cost, however, only takes the upload size into account and ignores the scaling done by the rigging. It also ignores the fact that LoD is usually effectively disabled for fitted mesh. That means, that if you upload the mesh at a very small size and with very low poly LoD models, you can trick the viewer into beiieving it is far less render heavy than it actually is.

The best way to estimate the actual render cost (or rather what the render cost would have been if the calculation had worked the way the Lindens intended), is to wear the item and get somebody using Firestorm to cam in close to it, inspect it and check the triangle count (you can't inspect items you wear yourself). Multiply that number by six and you have the base render cost and the rigged mesh multiplier. Then, if you want a more precise answer, add 384 for each 512x512 texture and 512 for each 1024x1024. (There are also a few other factors that may add to the render cost but the base render cost and the rigged mesh multiplier are usually the big ones.)

For example, if the head has 40,000 triangles and three 1024x1024 textures, the render cost would be at least 241,536. Compare that to the render cost figure the viewer gives you.

And now I'm shocked myself. This is the first time I've actually done the math and even though I knew it was bad, this is far worse than I expected.

[Selene Gregoire]

22 minutes ago, ChinRey said:

The "cheat" I was talking about, is one of the three fitmesh LoD bugs. A fitted mesh is of course automatically scaled to fit the avatar. So the size you upload it at doesn't matter to its looks or actual render cost. You can make it as small or big as you like, it's all the same.

The calculated render cost, however, only takes the upload size into account and ignores the scaling done by the rigging. It also ignores the fact that LoD is usually effectively disabled for fitted mesh. That means, that if you upload the mesh at a very small size and with very low poly LoD models, you can trick the viewer into beiieving it is far less render heavy than it actually is.

The best way to estimate the actual render cost (or rather what the render cost would have been if the calculation had worked the way the Lindens intended), is to wear the item and get somebody using Firestorm to cam in close to it, inspect it and check the triangle count (you can't inspect items you wear yourself). Multiply that number by six and you have the base render cost and the rigged mesh multiplier. Then, if you want a more precise answer, add 384 for each 512x512 texture and 512 for each 1024x1024. (There are also a few other factors that may add to the render cost but the base render cost and the rigged mesh multiplier are usually the big ones.)

For example, if the head has 40,000 triangles and three 1024x1024 textures, the render cost would be at least 241,536. Compare that to the render cost figure the viewer gives you.

And now I'm shocked myself. This is the first time I've actually done the math and even though I knew it was bad, this is far worse than I expected.

Not really what I was asking but ok. I think I'll just stick with what I have since it's not hard to keep the complexity below 50k as long as I don't try to wear a lot of jewelry. More often than not my complexity is under 25k with just the body, hands, feet, head, hair, a top, jeans and footwear on.

[steeljane42]

57 minutes ago, ChinRey said:

The "cheat" I was talking about, is one of the three fitmesh LoD bugs. A fitted mesh is of course automatically scaled to fit the avatar. So the size you upload it at doesn't matter to its looks or actual render cost. You can make it as small or big as you like, it's all the same.

The calculated render cost, however, only takes the upload size into account and ignores the scaling done by the rigging. It also ignores the fact that LoD is usually effectively disabled for fitted mesh. That means, that if you upload the mesh at a very small size and with very low poly LoD models, you can trick the viewer into beiieving it is far less render heavy than it actually is.

The best way to estimate the actual render cost (or rather what the render cost would have been if the calculation had worked the way the Lindens intended), is to wear the item and get somebody using Firestorm to cam in close to it, inspect it and check the triangle count (you can't inspect items you wear yourself). Multiply that number by six and you have the base render cost and the rigged mesh multiplier. Then, if you want a more precise answer, add 384 for each 512x512 texture and 512 for each 1024x1024. (There are also a few other factors that may add to the render cost but the base render cost and the rigged mesh multiplier are usually the big ones.)

For example, if the head has 40,000 triangles and three 1024x1024 textures, the render cost would be at least 241,536. Compare that to the render cost figure the viewer gives you.

And now I'm shocked myself. This is the first time I've actually done the math and even though I knew it was bad, this is far worse than I expected.

If you use firestorm you can inspect your own wearables through area search, just check attachments to show up in search and put your username into owner tab.

P. S.

At least bento heads made it better. Feel free to demo older non bento catwa heads (like "destiny", for example), they are over 1.2m triangles, so if go with your x6 formula they are around 7.5m of complexity. Those are still the only attachment to this day that made my old PC go to single digit fps on "inspection" (with highlight lines on at least), even belleza bodies weren't as bad.

Edited August 12, 2019 by steeljane42

[Da5id Weatherwax]

All this back and forth about the subjective "quality" of optimized content versus builds/avatars where optimization has not been done misses a fundamental point.

And that point is that a straight visual evaluation is subjective. It's also misleading.

Folks are decrying optimization as "looking worse" - when the "poorer appearance" is just as likely to be disliking the creators chosen aesthetic as it is a consequence of poor or overzealous optimization. It's perfectly OK to prefer one creators style over another's - that variety is something that makes SL what it is. But both could be equally optimized - you'd still prefer the one over the other but it would perform better!

Textures: At the closest normal range to view an avatar - say a fairly close-up portrait - or viewing a non-avatar build from a similar distance, any texture resolution where you've got multiple texture pixels in a single screen pixel is too high. It can be reduced without changing what appears on your screen at all. The appearance will not change in the tiniest amount from making this optimization. Your screen cannot display any higher resolution, so feeding it with a texture that has that higher resolution is a waste of VRAM and a cause of lag. This is not even a "best efficiency" optimization, that would be to set the cutoff limit where 1 texture pixel = 1 screen pixel at the most common distance to view the avatar or object - which would result in textures that were smaller still and unlike the "closest normal distance" optimization you would notice it in close-up shots. Nonetheless, it's still more optimization than most SL creators do. Most SL avatars and objects you'd have to have your camera pressed right up against the clipping plane to get 1 texture pixel to 1 screen pixel, if indeed you ever could. But when an optimization can be proven to make no difference to the displayed image - simply because of the display resolution compared to the resolution of the displayed texture - you can't claim it detracts from the quality at all. If you think it does you're simply wrong.

Polycount: Particularly for avatars, this one is more awkward. It is undeniable that where an avatars skin must deform, around joints for example, a higher vertex density and therefore a higher polycount makes that deformation smoother and more natural looking. But when you look at an arm and see that there's no discernible difference in vertex density between the areas around the shoulders, elbows, wrists and the areas in the middle of the upper arm and forearm then that's a badly unoptimized model. Period. Nobody is saying "You have to make your avatar look like crap as soon as it gets animated" - that isn't what optimization is about. It's about not putting in complexity that you actually don't need and will never see any visual improvement from. A higher vertex density around the joints is (mostly) fine and does give better, more natural-looking movement with the same animation. Way too many creators in SL just increase the vertex density everywhere to achieve this - it's easy to just add a subdivision factor to the model, after all - and then get all bent out of shape (ironically) when they get told their av is a lagmonster or they pitch a snit when somebody points out their creation looks solid in wireframe view, often claiming that it's that way because it's "high quality." This last is particularly pernicious because a user that has spent time and effort on their appearance will frequently pick this up and actually go shopping for these horrible things in the mistaken belief that it "looks better" And because they've told themselves that it does, then the Emperor's new doublet looks just fine.

[Kimmi Zehetbauer]

9 hours ago, Scylla Rhiadra said:

It's not just photographers, either: we know that there are people who will actually note on their profiles that they're not going to bother with you unless you have a high-end mesh body.

My man tried to get a DJ job at one club recently and was told since he didn't have a mesh avi, he was not welcomed.  So he wrote that club off. Sad...

[Scylla Rhiadra]

30 minutes ago, Kimmi Zehetbauer said:

My man tried to get a DJ job at one club recently and was told since he didn't have a mesh avi, he was not welcomed.  So he wrote that club off. Sad...

Sometimes it feels a bit as though SL does "shallow" really well.

[kiramanell]

23 hours ago, Scylla Rhiadra said:

This is a fairly decent looking avatar -- it strikes me that a big difference between this and the two others (which were definitely subpar compared to high quality SL avatars) is the texturing, which probably hides a multitude of sins.

So, I have questions. Questions questions questions.

-- How does the number of polygons or whatever we're using to measure this compare with the SL system avatar?

-- Alternately, how does it compare with, say, a Maitreya Lara and Catwa mesh head combo (which is probably a pretty common one)?

-- Does reducing ("optimizing"?) the number of triangles reduce the degree of customization you can introduce using sliders? Or is that just impacted by the avatar bones?

-- Does optimizing impact on the quality of the textures (i.e., skins) that you can apply?

So, this is what I can currently do in SL:

That particular outfit -- Slink body, Genus head, Vista hands, hair, dress, and boots -- currently costs me about 53,000 in complexity. How much would an avatar like the one you show degrade this kind of detail, and how much would I "save" if I used it?

ETA: I'm not sure how this all relates to "complexity," which I know is, in any case, a somewhat faulty metric, but in my experience it's my hair that tends to kick it up. With a simpler (but still good looking) hair than this one, I could do this same outfit for under 30,000 complexity.

 

*gasp* So beautiful! ❤️

Dat hair, what store is it from, please? I want quality hair like that too! My avi is currently at around 35,000 complexity, iirc, so there's room to crank it up a little. 😛 

[HeathcliffMontague]

On 8/12/2019 at 1:36 AM, ChinRey said:

Since bandwidth is essential too, we should also keep in mind that high speed connections simply aren't available everywhere, no matter how much you are willing and able to pay (and often no matter what your ISP promises).

Edit:

Even if you have a fast local connection, unless you happen to live close to the SL servers, the data has a long way to travel with many bottlenecks on the way. I just checked my connection to Tucson, AZ, the Ookla server closest to the SL servers and got 4.98 Mbps download speed. That's about half of what it was las time I checked and there's nothing neither I nor LL can do about it. (Come to think of it, last time I checked, USA still had net neutrality, I wonder if there's some connection there...)

Edit 2:

Yep, definitely net neutrality or some other US issue. I tried a Candaian Ookla sever and got 16.59 Mbps. And I tried Tucson again and got 3.53. Not only is international connection from the west of the USA slow, it's also very unstable.

You're in Scandinavia too, right? I have 100 Mbps and usually get at least 70 Mbps within Europe. Lately my Tucson tests have varied a lot. From 15  to as high as 60, depending on the server. Right now my lag is pretty low, but it has been really bad for the last few days.  Ping is always bad, close to 200 and sometimes even way beyond. But that is to be expected in Europe, apparently. What's new is packet loss. Whenever I'm lagging (and it can be really bad, even at high FPS) it's always that. Never had packet loss problems before.

Tracert used to reveal a problem just before Madrid, after I moved and switched ISP it's London. But the main problem seems to be close to LL in the U.S.

[ChinRey]

3 hours ago, HeathcliffMontague said:

You're in Scandinavia too, right?

Yes.

 

3 hours ago, HeathcliffMontague said:

Ping is always bad, close to 200 and sometimes even way beyond.

SL doesn't use dynamic routing so there's no chance to get it lower than that when you live as far away as we do. Some are much worse off. Interestingly osgrid.org does use dynamic routing and although their server aren't that much clsoer to us (they're in Canada), the ping time is considerably lower.

The bit ping time peaks (there's been a lot of that recently) are usually caused by the sim servers.

 

3 hours ago, HeathcliffMontague said:

Lately my Tucson tests have varied a lot. From 15  to as high as 60, depending on the server.

I thoguht maybe it was the Ookla server in Tucson that was acting up but I also found a non-Ookla server in San Francisco and it was pretty much the same. The other locations I tested at the same time were the town closest to where I live (exactly where is classified info), Trondheim, London, Dublin, New York, Houston and Montreal. Most servers were Ookla but not all. Houston was a little bit slower but not nearly as slow as Tucson and SF. There were only minor differences between the other six.