Specs of the machine used for testing:
CPU i7 4790K @ 4.3 GHz
RAM 32 GB DDR3
Network 1 Gbps (300 Mbps guaranteed, unmetered)
HDD Samsung 850 EVO 120 GB SSD
OS Windows Server 2012 R2 Standard (x64, of course)
CPU-Z validator
This is not server-grade hardware. However I went for the i7 because of its very good single-thread performance. And it was cheaper (25€/month thanks to a promotion) than any server with a comparable Xeon CPU. An i7 is probably fine for game servers anyway since features like ECC aren't really all that important.
General Notes:
- Vehicles are much heavier on the CPU and network than players
- DLC maps will always load faster
- Loading times will be much worse on mechanical drives
- Bandwidth and CPU requirements on high frequency are much higher than on 30 Hz
- I set the affinity of each instance to two of the CPUs 8 threads. That way they don't interfere which each other and reading the CPU load of each individual instance becomes easier
Those might be boring to most of you but I found it interesting.
Startup process (average):
Vanilla Maps: 26s
DLC Maps: 21s
Level Load Time (average):
Vanilla map: 12s
DLC Map: 8s
Idle Load:
150-200 MB RAM (depends on loaded map, etc.)
0.4% CPU
Extreme Scenarios:
Those numbers were recorded over short amounts of time in a round and represent performance around peak loads.
120 Hz with Vehicles
Map: XP1_004 (Wake Island), Assault
Players: 32/32 + 1 Spectator
Vehicles enabled
CPU:
~16% total for vu.exe
90-100% of one core (main simulation thread)
+ 5-10% of another core (other stuff)
Memory: 385 MB
Network:
Outgoing:
min 13 Mbps
max 23 Mbps (!)
avg 18 Mbps
Incoming:
min 2 Mbps
max 3 Mbps
avg 2.6 Mbps
Server FPS:
Note: Everything < 120 will lag at 120 Hz, < 130 is already bad but playable
min: 125
max: 169
avg: 150
60 Hz with Vehicles
Map: XP1_002 (Gulf of Oman), Conquest Large
Players: 62/64 + 0 Spectators
Vehicles enabled
CPU:
~18% total for vu.exe
90-100% of one core (main simulation thread)
+ 10-20% of another core (other stuff)
Network:
Outgoing:
min 18 Mbps
max 28 Mbps (!)
avg 22 Mbps (!)
Incoming:
min 3 Mbps
max 4 Mbps
avg 3.5 Mbps
Server FPS:
Note: Everything < 60 will lag at 60 Hz, < 70 is already bad but playable
min: 52
max: 67
avg: 55
Players noticed and mentioned rubberbanding in chat. Once FPS were above 70 again those issues were gone.
120 Hz Infantry only
Map: MP_017 (Norshar Canals), TDM
Players: 32/32 + 0 Spectators
No vehicles on that map.
CPU:
~14% total for vu.exe
80-90% of one core (main simulation thread)
+ 5-10% of another core (other stuff)
Network:
Outgoing:
min 5 Mbps
max 9 Mbps
avg 7 Mbps
Incoming:
min 1.5 Mbps
max 3 Mbps
avg 2 Mbps
Server FPS:
Note: Everything < 120 will lag at 120 Hz, < 130 is already bad but playable
min: 184
max: 243
avg: 196
There is no data on 30 Hz available from my side. But its safe to assume that it will run on just about anything.
Conclusion:
In general it came as a surprise to me how high the bandwidth demand of high frequency really is. Those of you that have servers with metered connections might want to make sure that they have enough traffic per month available to run high frequrency servers.
Most server CPUs are not overclockable so you're stuck with whatever performance it has.
Most newer Intel Xeon Quad Cores should be good enough for 32 player 120 Hz (without vehicles or maps with few vehiles), 24 player 120 Hz with vehicles or 64 player 60 Hz Infantry only or 32/48 player 60 Hz with vehicles.
At this point only a Xeon 1245v2 and a Xeon 1240v3 have been tested by me. Both did fine with 48 players at 60 Hz and 24 players at 120 Hz. 32 players at 120 Hz (with vehicles) was only tested on the 1240v3 and was not playable.
I might edit this in the future when I get some more readings. But those are the best I have right now.
