Competition from IBM

[zero]

Cell Broadband Engine Project Aims to Supercharge IBM Mainframe for Virtual Worlds - or is it something totally different?

[ChrisM]

Totally different.  What IBM is doing is marrying a Cell based server platform with a, wait for it....mainframe.  Aside from the obvious (i.e. single points of failure, physical limitations due to hardware stacks, etc.) there are other questions like how much is it going to cost to buy hardware this big? How complex is it going to be to build for both Cell and mainframe systems?  Who in the games industry has expertise in building or managing a solution like this?  Their name for this new system? "Gameframe"

To quote a very good friend of mine in the games biz, in an IM discussion last week, "...well i'm sure they'll sell pretty much zero of those...  sounds like typical ibm not really understanding what they are doing."

-Chris

[kevglass]

So same target, different approach that you'd see as less appropriate than DarkStar?

Kev

Games published by our own members! Check 'em out!

[ChrisM]

So same target, different approach that you'd see as less appropriate than DarkStar?

Kev

Same target, VERY different approach.  This is similar to the fantastic failure of IBM/Butterfly.Net GRID "experiment" done 3+ years ago.  Here is the fundamental difference between the approach that IBM and Sun are taking in the space.  IBM's answer to online game technologies is to provide tools that a game developer can use to create their own build a server solution.    None of what IBM offers actually helps game developers understand HOW to build such a beast.  For that, you must rely upon IBM's "services" division.  The approach we took with Darkstar is pretty much the opposite: present a highly scalable, fault tolerant, multi threaded system as a mono-threaded server accessible through APIs.  Our goal is to provide a solution, IBM's is a jumble of tools. 

Take a look at how IBM breaks down their game server architecture:



Notice anything?  Completely hardwalled architecture.  No scaling across the entirety of the service, no resource allocation between their "zones", etc.  On top of that, the new IBM "Gameframe" solution is a mix of mainframe and Cell based servers.  How proprietary is THAT?!?  Consider that the cheapest IBM mainframe, the z9, starts at $100,000 and no pricing on those Cell blades, and you are talking about Sony/Blizzard prices infrastructures.  With Darkstar, you can start on your single $500 dev box and scale up from there.

So again, same target, and IBM is using the same old approach that the industry is trying to move off of.  Not only do I believe it's less appropriate, I believe they are moving backwards.

-Chris

EDIT:  Another example of more exotic hardware being used to smooth out implementation issues is the solution for EVE-Online.  In order to get to the 30,000 players per shard number that they claim, they had to install military grade Solid State Disks that can handle 400,000+ random I/O per second.  An this is in a game that is instanced.  Again, throwing more hardware at a software problem is best left for last generation thought of online game technologies and for Vista (oooohhhhh!!!!).

[blahblahblahh]

EDIT:  Another example of more exotic hardware being used to smooth out implementation issues is the solution for EVE-Online.  In order to get to the 30,000 players per shard number that they claim, they had to install military grade Solid State Disks that can handle 400,000+ random I/O per second.  An this is in a game that is instanced.  Again, throwing more hardware at a software problem is best left for last generation thought of online game technologies and for Vista (oooohhhhh!!!!).

Sorry ... Eve is "instanced" ? Where did you get that from? Eve is usually known as the "one server, no shards, everyone in the same massive virtual world" game. Obviously, they don't have everyone talking to everyone all the time in one area where they can all see each other, but using that as a way of compartmentalizing the load is surely the same as you're trying-for with darkstar?

If anything, I would have thought Eve was an example of exactly what DS can achieve - no server goes unused - it's just that they're already at the "sony/blizzard" scale you were talking about, so they're already running on big iron?

Since you're doing a comparison here, could you outline the differences between butterfly and darkstar? On the face of it, they appear to be very similar architectures - without looking at the specifics, DS always reminded me of a standard grid application, with the transparent distribution etc. I think it would be very handy to know how DS improves on that (for the record, I too was never a fan of butterfly )

[Jeff]

Sorry ... Eve is "instanced" ? Where did you get that from?

From me.  I got it from the EVE guys themselves at a talk at GDC>

Chris is fundamentally correct.  Its not *quite* fully isolated instances, but close.

They take advantage of the fact that their simulation is fundamentally deterministic (newton is newton) and run separate instances of the simulation on each server with very limited inter-communication.

Ofcourse this breaks down in combat which is not so deterministic.  they admitted in Q&A that combat alone takes 35% of their compute resources currently and that it is scaling very badly.

This is *nothing* like Darkstar, which does a true shared single instance of the world across all the nodes.