SL and Server Architectures

Dan Terdiman has a story over at CNet on Second Life's server architecture, and the way in which servers are loaded generally in the MMOG arena. References to our Cory Ondrejka and Mike Sellers, and some interesting questions about balancing and server architecture design.

Comments

[nn]

1.

Daniel James' followup to this article.

[greglas]

2.

That Daniel James post is good stuff...

[Mike Sellers]

3.

Daniel's post is well worth reading. Personal blogs are just the sort of place to expand your thoughts when an article necessarily boils down a 30 minute conversation to one quote. Alas, I'm too lazy for that and have only TN.

As Daniel says,

I will go on the record that 3-20 concurrent users per CPU is not reasonable for a service that hopes to scale to become the defacto Metaverse. To run with 500k concurrent users in China, like Audition, you’ll need all the servers in Shanghai, and more.

This is one of the central issues that I see with SL -- and again, I'm not saying this because I want to be the jackass that helps to kick down their barn, but because I'm concerned that the combination of their strong potential, incredible PR machine, and Icarus-like path could hurt the long-term acceptance and innovation of online worlds.

Call it the beginning of what could become, if we're not careful, the Metaverse Bubble: a few years ago online worlds were the exclusive province of a few strange geeks. We dreamed idly of the day when we'd be accepted; we said things like, "yeah, there are a ton of PhDs just waiting to be had here, but we get no respect from academics" and "someday, even your mom will play one of these games." And now we have increasing academic respect. Many moms are playing/inhabiting online worlds. And the level of social (and financial) currency that follows such developments is rising too.

We have the opportunity to really do well here, and to do some good: to do more than make the equivalent of "He-Man Online" but to create the first truly new form of entertainment (and thus education) in centuries. This could be the start of a golden age for online worlds that leads to things we can't even conceive yet. Or, if we have a few too many failures -- if in the next year the frighteningly large crop of me-too fantasy MMOGs wastes away in the deep shadow of World of Warcraft and Second Life's technical and/or business model stalls despite the company's brave front -- well, we could all be shoved back a ways. The naysayers would have a field day. Instead of the new golf, we'd be the old Rubik's Cube: "See, it was just a geeky thing all along; no one I know played those games." The academic, popular, and financial support could dry up like Ross Perot's political base.

Now, I'm not saying that Second Life is the harbinger of the end of online worlds. I hope they continue to grow and increase engagement in online worlds with those not yet familiar with them. But their technical and business model coupled with their magazine-cover PR makes alarms go off in my head. Despite Philip Rosedale's analogy, SL isn't "just like Google" -- the usage and CPU/RAM/bandwidth profiles for any online world are completely different from that of a search engine. Maybe, as Daniel says, these guys are just smarter than the rest of us and Cory will win his bet, looking out over a vast sea of SL residents (and an equally vast sea of CPUs, sucking down the West Coast power grid), in a future where WoW has become a sleepy backwater town. Or maybe LL will nimbly avoid this issue by surmounting the obstacles to a shared-but-distributed world, where every resident's computer is also potentially a server (a problem on the order of creating an in-home fusion reactor, IMO). Stranger things have happened. But not many.

[Larry Kyrala]

4.

It's a shame that the CNET article didn't go into more detail about the actual architecture. Saying that it is simply more scalable that Blizzards and that "it works for Google" sounds too much like marketing hype.

Google's architecture works because incoming queries are independant of each other. In other words, they mesh perfectly with the highly scalable principles of a REST architecture. (Query times can remain roughly constant simply by scaling the number of CPUs.) However, MMO's don't share that property. The whole point of being in an MMO is to interact with other people and the environment. User interactions scale exponentially with respect to the number of users interacting regardless of how many CPUs you use.

WoW has a fixed number of zones, plus some filtering based on the maximum interaction range. SL seems to have a configurable number of zones that can be added to on-the-fly. This allows them to scale their world space "infinitely" as they claim. However, if a large number of people in SL decide to congregate in the same physical location what happens? Is it like the auction house in Iron Forge all over again?

There are also some novel approaches to the problem coming from the grid computing area that focus on optimizing the performance of the interactions rather than partioning them. IBM did some interesting research that allowed 200+ people to play in an "unpartitioned" Quake server instance using their grid computing platform. A company they partnered with, Butterfly.net sounded promising at first, but apparently didn't make it out of the gate... grid technology is still pretty new.

[kathygnome]

5.

"Is it like the auction house in Iron Forge all over again?"

No, that's what it's like all the time. When a big group gathers its far worse.

[Young Freud]

6.

"Is it like the auction house in Iron Forge all over again?"

No, that's what it's like all the time. When a big group gathers its far worse.

Yes. I've witnessed bizarre effects where over 40 AVs would overload a sim and crash it. There was ways around it, like hosting events at the corners of 4 sims to increase event attendance, but usually at 30+ players, sim reliable became a bit unstable, especially if they where all blinging out and running scripts, let alone doing anything physical.

Recent releases have increased it significantly, IIRC, to somewhere 80-100 per 256mx256m sim. However, even with this, I do think a grid-based, manufactured-content MMORPG might be an interesting concept, especially with an expanding narrative.

[Prokofy Neva]

7.

If there are only ever 2 people on a sim (depending on the class of server, it will hold 2-4 sims in theory) then, is that a poor use of space or CPU power? Or is that 2 people who bought a server for as much as $1500 USD for the upfront purchase price, and then pay $195 each month for the maintenance fee, except that it is empty except for them, when they want it? The land=server formula is the formula that enables people to spend this kind of money - if you asked them, "How would you like to spend $1500 US up front to be on the Internet and have server space, and then pay $195 ever after," they'd say no, or spend only $32. But call it "land" and make it an immersive world, and people will pay, and spend lots of time on it.

I'm wondering how you can keep the value of "virtual" estate and not have that 1:1 relationship of "land" and "server space".

If you look out over RL America, the 'green dots' aren't distributed so efficiently there, either.

[Cruise]

8.

bgdd