Scalability and Memory Management for Many Dynamic Objects in Large Space
Members - Reputation: 546
Posted 01 October 2012 - 01:47 PM
Here's a description of the system as I envision it: The C++ engine will serve as a means of communication between Lua/OpenGL and Lua/SQLite in addition to other less relevant things like file loading. Game objects (that is, things that can be interacted with directly) will be handled Lua-side, communicating with each other through a messaging and event system routed through C++. For its part, C++ has absolutely no idea what an object is or does, it just blindly passes messages. These Lua-side instances will be divded (by space) into an active group - those objects that are on the screen - and an inactive group. When an object moves from active to inactive, it is removed from Lua (and thus RAM) and ported to SQL. Every 2~7 seconds, these objects are pulled back into Lua, updated, then sent back to SQL. Objects in the active group are rendered, undergo physics simulations, ect. To the player, the whole world is doing this, but in actuality it's just approximating what might have happened (basic snap-to pathfinding on an extremely simplified navmesh, for instance). Quitting and saving is merely a process of storing all remaining active objects in SQL. Loading just requires a query to locate objects in the same or neighboring spacial partition as the player/viewpoint.
The reason I ask about this is because one, I suspect I'll be dealing with hundreds of thousands of objects spread out over hundreds of kilometers of space. They're also quite data heavy, and I suspect processing each object will take quite a few ms to complete (a lot of object lookups/queries, a lot of unfortunately redundant checks). I can also see a bottleneck arising. The C++ engine is only aware of Lua-side objects. If an object is SQL-side, when C++ attempts to send it a message, Lua will have to check if the object exists, find that it doesn't, pull the object from SQL, process the message, and then port the object back to SQL. For every message. Something tells me this is too slow for interactive frame rates, but I have yet to test the system. Ideally, I only want to use SQL as a save state and to cut it out of the play-process entirely. I chose SQL because it's fast, SQLite is, well, lightweight, SQL works well for random access lookups, and I'm familiar with it. This is why I phrased this as memory management, since I can't see thousands of lua objects being friendly to my 4GB total RAM budget.
I'm targeting systems with 1GB VRAM, 4GB RAM, and 4 processor cores. As such, I also want to take advantage of multi-threading to make use of those cores. Typically, I avoid multithreading like the plague, or only allow each thread to operate on completely different sets of data. For this game, I wanted to implement a task system, where the main thread can create tasks for a pool of 3 identical subordinate threads to process (I suppose a configuration file would let the user select more or less to match their core count, but the goal is 4 well utilized threads). Thus, individual updates of offscreen and therefore insiginificant objects can happen on a separate thread (right along with LOD file management), so all the main thread needs to concern itself with is physics and rendering. When the game is loading, for instance, the subordinate threads are doing the heavy lifting while the main thread renders the now loading video/progress bar/ect.
My questions ultimately boil down to these: is there anything fundamentally wrong with this approach? Is there a more optimal approach, or a simpler approach that will fully make use of the target system? Can you see potential hazards in this approach that can be avoided through careful design?
Members - Reputation: 604
Posted 02 October 2012 - 07:37 AM
and to minimize streaming and stuff, divide your world into segments which cointans a set of objects. (quad tree) that would probably speed up things alot.
Edited by Tordin, 02 October 2012 - 07:37 AM.