AI and Server/Client Performance (Test Results)

I did a test with BigMorgan regarding AI and Server/Client Performance. I think you guys will find these results fairly interesting.

I'll start off with my Server Specs, Server Config and both Client Specs.

Server Specs

Xeon Quad @ 2.66ghz

4GB DDR3

100mbit+ Upload

Server Config (MaxSizeNonGuaranteed is commented out)

MaxMsgSend=8192;

MaxSizeGuaranteed=32;

MinBandwidth=2000000;

MaxBandwidth=2200000;

MinErrorToSend=0.001;

See this thread for Server Config Discussion

Client #1 Specs

Q6600 @ 3.00ghz (OC'ed)

4GB DDR2

NVIDIA GTX 260 Core 216 896mb VRAM (Stock)

16mbps/1mbps Connection

Windows 7 x64

Client #1 Specs

i7 920 @ 2.66ghz (stock)

NVIDIA 9800GTX+ 512mb VRAM (Stock)

6GB DDR3

20mbps/3mbps Connection

Windows Vista x64

Here's how we tested. We used my latest warfare, set a unit build time to 0 seconds and give ourselves endless money. We were on opposite teams and just sat there as units built. Keep in mind these units are local to the client and NOT the server. Now into some of the results.

We had a solid 45-50fps on the server until around 130 units and then it began to rapidly decrease till it got down to about 20fps on the server at about 160 units. At around 200+ units, the server FPS was hovering around 7-11fps.

I remember somebody saying somewhere on these forums that if AI were local to the client that they have near to zero effect on server performance. We've sat here and proved that VERY wrong. An interesting thing we also noticed is that AI local to another player could cause lower performance on the other client. I let client #2 build 100 AI and I sat there with 0 AI, but my client FPS steadily decreased. He and his AI were 7000m away from me on the map.

We're going to do more testing over the weekend, but I thought I'd post my findings here so people can try to better understand how AI will affect both client and server performance greatly.

Much appreciated that you're doing this.

Quick question on the testing method. Does a mission like Warfare not have significant overhead? Even if not in its own code what about the tasks it assigns the AI? It's common knowledge at least among server admins and mission makers that waypoints like DISMISS are a lot hungrier than non-moving AI or regular MOVE AI.

The warfare was semi-stripped. No looping scripts (the ones that take the resources) were running.

Only the unit buy function.

I'll try to setup a mission that everyone can use to test.