9 replies to this topic

[xiongyouyi]

Time-slicing and multithreading for A-star pathfinding, witch one is the best?

[alexjc]

What are your requirements?

In many situations you should need neither for the following reasons :-)

• Time-slicing implies your pathfinding requests take longer than a frame. Best optimize them or use an algorithm that fits comfortably within a frame. Then do an integer number of them each frame.
• Multi-threaded pathfinding is used on PS3, but often on XBox360 or PC there's a separate (single) thread for pathfinding.

Keep in mind that either of these choices can result in your AI being much less deterministic if you're not very careful!

Alex

[xiongyouyi]

What are your requirements?

In many situations you should need neither for the following reasons :-)

• Time-slicing implies your pathfinding requests take longer than a frame. Best optimize them or use an algorithm that fits comfortably within a frame. Then do an integer number of them each frame.
• Multi-threaded pathfinding is used on PS3, but often on XBox360 or PC there's a separate (single) thread for pathfinding.

Keep in mind that either of these choices can result in your AI being much less deterministic if you're not very careful!

Alex

Thanks Alex! I'm planning to make a RTS game, use NavMesh and RVO for AI navigation, so if there are hundreds of units request pathfinding at same time, I don't know which solution is the best.

[ApochPiQ]

If you have to service large volumes of path requests, just do them in batches - 20 in one frame, 20 in the next frame, and so on. (20 is just a made-up number, of course.)

You shouldn't really ever run into a situation where you absolutely have to have hundreds of path results in the same frame.

[xiongyouyi]

If you have to service large volumes of path requests, just do them in batches - 20 in one frame, 20 in the next frame, and so on. (20 is just a made-up number, of course.)

You shouldn't really ever run into a situation where you absolutely have to have hundreds of path results in the same frame.

I want to test multithreaded path finding for large units.

[ApochPiQ]

What do you mean by "multithreaded pathfinding" though? Running multiple path queries on separate threads concurrently, or using multiple threads for a single query?

[xiongyouyi]

What do you mean by "multithreaded pathfinding" though? Running multiple path queries on separate threads concurrently, or using multiple threads for a single query?

I mean running multiple path queries on separate threads concurrently, it will not affect the fps.

[ApochPiQ]

Assuming you use a fairly standard approach of high-level pathfinding combined with local steering for avoiding other units, you should be able to just run a copy of the path search on as many threads as you like (assuming your searches don't mutate any state of course).

[Ashaman73]

I use a combination of both, multiple threads calculates multiple pathes concurrently in a timesliced manner. The reason is, that my waypoint graph will be manipulated dynamically, by either setting waypoint attributes (open/closed doors etc.) or by changing the topology of the whole graph.

[wodinoneeye]

If you have multiple pathsfindings to do, run that process on a thread with affinity set (especially if most of the map can fit in its cache) let each one finish before starting the next (especially if the map is dynamic -- units positions count as 'terrain' block movement factor) so that that result can be passed on to another process to get something done..

Multiple core - definitely have as a seperate thread (and if you profile and find that it gets done and has time left on that core you might have someother lightweight task for it to do to fill out its capacity.


If you have too much pathfinding to do then you need a priority scheme to figure out the best ones to calculate first (and do less importants ones less often - but still get done within some maximum timespan)
Eaxmple would be units who's target has moved over one that hasnt....

See if you can do a hierarchical scheme to simplify the processing (depends on terrain but interiors with obvious 'portals and coarse paths limited by the terrain are good candidates for this)

Also recycling the old path (if you have memory to keep it) can shortcut processing on subsequent pathfindings.

Optimizing the pathfinder itself can often speed up the processing by a magnitude (make use of the specifics of your map data, instead of a generalized A* )