267 CPU/GPU performance
Martigen
Australia Join Date: 2002-11-01 Member: 2714Members, NS1 Playtester, Contributor, Reinforced - Onos
Australia Join Date: 2002-11-01 Member: 2714Members, NS1 Playtester, Contributor, Reinforced - Onos
Firstly, 267 feels much smoother and it's a joy to play. The more I play it the more I'm impressed.
Secondly, for what it's worth here's a screenie of CPU and GPU usage during a game on Woozas (leave your small mind at the door, please) with 32 players, end-game, Mineshaft, with an assault on Deposit. I'd say roughly 3/4 of the players were present (people in the spawn queue, etc).
Relevant info:
System: Res - 2560x1600, CPU - 3970X @ 4.5, GPU - 3xGTX680 4GB, Storage - SSDs.
NS2: Maximum details, except Ambient Occlusion disabled. Texture Manager set to 2GB+. Physics Multithreading on. DX9.
I didn't record FPS over time (though I may do this next) as initially I had this open to watch effect of the physics multithreading, but it was largely capped at 60fps (going by 'fps' console in NS2) except for occasional drops. Woozas currently has connection problems with ISP, so hard to say if these are client side++, server, or connection at the moment. The screenie is a guide for resource usage.
++ In the past I've had hitching even while CPU and GPU usage remains < 99%. 267 seems to have reduced this.
Interesting to note VRAM usage peaks, but remember this is 1600p. SLI Scaling is excellent. CPU usage is also distributed well, with just a few peaks hitting single-threaded max on core 1 at end-game.
Unrelated, but for those wondering about temps and power level: first two cards are water-cooled, and of different brand than the third.
Secondly, for what it's worth here's a screenie of CPU and GPU usage during a game on Woozas (leave your small mind at the door, please) with 32 players, end-game, Mineshaft, with an assault on Deposit. I'd say roughly 3/4 of the players were present (people in the spawn queue, etc).
Relevant info:
System: Res - 2560x1600, CPU - 3970X @ 4.5, GPU - 3xGTX680 4GB, Storage - SSDs.
NS2: Maximum details, except Ambient Occlusion disabled. Texture Manager set to 2GB+. Physics Multithreading on. DX9.
I didn't record FPS over time (though I may do this next) as initially I had this open to watch effect of the physics multithreading, but it was largely capped at 60fps (going by 'fps' console in NS2) except for occasional drops. Woozas currently has connection problems with ISP, so hard to say if these are client side++, server, or connection at the moment. The screenie is a guide for resource usage.
++ In the past I've had hitching even while CPU and GPU usage remains < 99%. 267 seems to have reduced this.
Interesting to note VRAM usage peaks, but remember this is 1600p. SLI Scaling is excellent. CPU usage is also distributed well, with just a few peaks hitting single-threaded max on core 1 at end-game.
Unrelated, but for those wondering about temps and power level: first two cards are water-cooled, and of different brand than the third.
Comments
Because: It would be small minded to suggest that the client doesn't experience a loss in performance due to larger, and unsupported, entity counts.
And as noted, it's a testament to the engine and the work in 267 to increase performance, specifically late-game. The CDT asked us to test the new multi-threaded physics and report back, so I fired up Windows task manager and Inspector's GPU graphs for a game, and was impressed. I hope I'm not wrong, but I thought the CDT might appreciate seeing how it scales on a 6-core when pushed to high loads.
That's nonsensical. They show how the engine is scaling for that given load, regardless if it's 'supported' or not. They are what they are.
I'd like to think the CDT would find it informative as a use case. Running software beyond spec can help reveal issues that can impact when it's run within spec.
I think that the few servers that support the very large player count is fine for some players, and I enjoy it once in a while as well, and if the recent patch makes both the main servers and also the few large player count servers happy, then that's great.
I agree that testing far ends of the spectrum - even if the conditions are beyond a practical scope - is typically a good idea to see where we can expect things to break.
But , in this particular instance what we need more is your specs (which you provided) and comparison pics of multithreaded turned on and off in a similar (ideally the same) late game scenario that is within the supported and intended limitations (24 players or less). You can do so with hand made graphs, or simply submit plogs. (instructions in 267 thread)
This is to asses whether it's worth the risk of potential bugs so that maybe one day we can default it to on for all users, provided the data you share demonstrates worthy results of course.
The hitching reduction is undeniable, though I've had several people comment on how the overall FPS appears to have taken a hit. Hard to say whether this is an actual overall fps reduction or due to some other kind of 'hitchy' behaviour (or if its just people's imagination )
@Martigen
I would also like to see a frametime graph for your rig. And if you can, try one with ambient occlusion on. I also have Tri-SLI and get amazing scaling in NS2 (partially due to my insistence with Nvidia to standardize and implement proper SLI bits which I researched!).
Check out my thread here (GPU BOUND Hitching (frametime disparity w/ higher GPU load)): I am documenting the last remaining hitches in NS2 and it would be great if you could contribute similar data.
It also heavily depends on the player demands. I remember so many moments where we were sitting in front of some computer playing some game and it was running like crap. I've asked if really low framerate is bothering anyon, because it was hurting my eyes. And it was ok to the rest. I was the only one noticing it and being annoyed by it.
Same is here. I can sense when mouse motion changes from absolutely fluid (start games) into micro sluggish (mid and end game). And that's so horribly annoying it's still driving me nuts even with 267 update. It's like trying to hit a coin at 2km distance with a sniper rifle made out of jelly. It just feels wrong. 267 has slightly posponed this behavior, but not by much. It's the 80fps that feel like 25fps thing. Can't explain it any other way, because FPS counter is still showing really high framerate despite endgame slowdowns, but my eyes are registering motion that doesn't fit that framerate.
Didn't mean to sound harsh, we just badly need data that's within a specific parameter.