EVE Performance Issues (looking for video card suggestions)

So I've started to process to upgrade my ... well everything in support of my isboxer fleet in EVE and better performance for my vmware workloads.

Current system:
AMD Phenom X6 2.8GHz
16GB DDR3 1600
240GB Intel SSD
GTX680
GTX770

Both cards run two monitors. Each runs a 22" 1920x1080 and a 27.5" 1920x1200.
Been using the GTX680 on the 22" for full screen stacked config and the 2 27.5" for my dashboard views. The 4th monitor is for jabber and TS and OOG browser.

My system runs a 6 client fleet pretty well. Starts dragging pretty bad with 7 clients when there are a lot of other ships on grid. I have 11 clients in total and I want to use them all with good performance.

I loaded up all 11 to do some performance testing. My system is just wrecked by this. Even with 7 or 8 clients cpu is 95% or higher most of the time. I installed GPU-Z and found I was also using nearly all (probably all) of the 2GB of memory on my GTX680. Dropping all my clients but my main one to low (performance) settings lowered the GPU memory used by about 500mb but cpu was still an issue.

So that's history. I've ordered a new precision workstation from Dell (Precision 7910) with dual 8 core xeon processors (16 cores plus 16 virtual HT cores) and 32GB RAM along with other goodies. I opted for the default pos quadro video card because they don't offer GTX as a configuration. This solves the cpu bottleneck and probably any ram bottleneck if present. That just leaves the video card question.

My initial thought is get as much ram in a video card as I can to support the 11 clients stacked full screen on the main GPU. This makes me think of the GTX Titan Z or the Radeon 295x2. The question on these though is I read on the AMD card that if you don't run full screen you don't get the benefit of the second GPU and memory. Does that way isboxer stack full screen make the game really full screen or is it not true full screen? Usually i associate full screen with not being able to move the mouse off the main screen without hitting the windows key or alt tabbing to another app. Haven't found anything saying either way about gpu limitations on the Titan Z. Anyone have any input on the effectiveness of these high end cards for multiboxing?

[bob]

ISboxer forces a windowed mode, despite the game setting. Full Screen's proper name is Exclusive mode, and it is exactly that.

As for your other questions, Crossfire does not support windowed mode, so ATI crossfire is not good for mutliboxing, unless you are old school and ALT+TABing it, with a load of graphics resets between as each game switches in and out of exclusive mode..

NVidia does support SLI in windowed mode so it is the one to go for.
TitanZ's are tricky though, you really only get 6GB vram when running in SLI, plus they are really power hungry and have very specific requirements on the PSU, and less performant that a 980. You are also driving a shedload of pixels, and who knows how many monitors a TitanZ will natively drive at one time, and whether you can have monitors separately on each GPU (anyone with a TitanZ care to provide some input?).

Personally I'd look at the 970's or 980's. You get at least double the amount of vRam you already have and it sounds like that would be enough, they are cheaper (x3 of them), and less power hungry, and you can run them individually if you want , and they have some good grunt matching and exceeding the Titan.

My biggest concern is again not having enough memory on the card to support smooth gameplay. I would rather get something beyond what I need than spend the money on something that doesn't work.

If I can't get the 12GB out of a Titan Z maybe a 6GB Titan is the better way to go?

Does SLI support running more than one monitor now? I haven't used SLI in about 5 years. Will running two 980's provide 8GB of ram to the stacked full screen display? What about Tri-SLI? Does anyone have experience with that?

The tower I am getting supports 3 high end quadro cards and has 6 8pin pcie power connectors with a 1300 watt power supply so I'm sure it will run nearly any configuration I throw at it.

[MiRai]

My biggest concern is again not having enough memory on the card to support smooth gameplay. I would rather get something beyond what I need than spend the money on something that doesn't work.

People run more than 11 clients on 3-4GB of VRAM. Are you requiring that each game client be set to maximum quality?

Does SLI support running more than one monitor now?

SLI supports at least two monitors (outside of Surround) assuming you're plugging all the monitors into the "master" GPU.

I haven't used SLI in about 5 years. Will running two 980's provide 8GB of ram to the stacked full screen display?

SLI mirrors the memory between each GPU, which is why a Titan-Z marketed with 12GB only actually has 6GB.

What about Tri-SLI? Does anyone have experience with that?

What about it? The more GPUs you throw into SLI the more performance you can (theoretically) expect to get. You can (definitely) expect diminishing returns on each GPU you continue to add to the mix (maximum 4).

The tower I am getting supports 3 high end quadro cards and has 6 8pin pcie power connectors with a 1300 watt power supply so I'm sure it will run nearly any configuration I throw at it.

Does it specify which Quadro GPUs the board supports in 3-way SLI? A quick Google search for a Precision 7910 brings me to an owner's manual with this image as your potential motherboard, and looking at the layout leads me to believe that your can only use 3-way SLI if you had single slot GPUs (e.g. Quadro K4200). All gaming GPUs that are worth talking about require two slots at a minimum unless they're water-cooled, so I'd suspect you're only going to be running 2-way SLI at most.


Do you mind if I ask where you're getting your information from? What, or who, is telling you that you need a dual-CPU Xeon system with 8-12GB of video RAM to run 11 EVE clients?

I would like to run them on maximum quality settings yes. I run a mixed fleet with dps/logi/ewar/boosts and while my main screen has all that dashboarded some situations I switch to the others off my main.

I was referring to tri-sli in the sense that you would have even more processing power and memory, but since you said SLI mirrors the RAM that's not a viable option.

Since 3 card SLI isn't really worth it for my setup I don't have to worry about that but yes I was thinking of water cooling or potentially finding long sli cables. I remember a long time ago i found 12" sli connectors and that would work for the split slot setup the board has.

No one told me to get a precision. I'm a NetApp solutions architect and no stranger to understanding and measuring compute needs for datacenter infrastructure. I was originally going to go with dual 6 core as 1 client per core is an easy number but then thought since I'm spending so much I should put some of my virtual machines on this system as well so I opted to spend $400 more for dual 8 core processors.

I haven't done the hardcore gaming thing in a long time and I don't plan to so my lack of video card knowledge and performance are the things you guys are filling in the gaps for and i definitely appreciate it!

I'm still concerned about only having 4GB of memory especially since even if that worked it would probably be close to the ceiling and if for some reason I decided to add anymore clients I would be back in the same boat. I guess since I was planning to buy a titan z that buying a 980 to pair with my 680 is a start and if I need more I could move up to a titan black/980 pair and still be at the same cost as a titan z would have been.

[bob]

I'm a NetApp solutions architect

Should I blame you then?
Whomever keeps suggesting the Big fucking Raid6 for everything should be shot.


But lets forget about that for now. We here to talk about serious stuff.

With SLI, you can only pair like GPU's, so 980 with 980, or Titan with Titan. Preferably same memory size, and memory/GPU speeds too, and if being pedantic, the same brand/model. The closer you can get to twins the better.

Yes, Tri SLI would give you more compute power, but overall there is still only 4GB of vRAM; that might be enough anyway. Have you checked to see how much VRAM you need to run 5 clients at maximum? Your first post indicated that you could run 6 clients well, but this didn't provide any numbers definitive, values. You did indicate that your CPU started to struggle at 7 clients, but no mention of the vRAM maxing out until you were running 11 clients, at what settings? Those suitable for you to play?

I suggest starting off getting some potential idea of what you may need. You should probably do some measurements for 1, 2, 3, 4, 5, 6 clients until your vRAM hits it's limit, so you can get an idea of the increasing demand for each client, then you are in a far better position to size your needs. You should be focusing specifically on vRAM requirements for this.
You could do the same for GPU, but that is harder to quantify as the CPU can come into the numbers when it starts being the bottleneck. You can only go up to the ceiling where the CPU starts to peg out, after that GPU utilisation is variable. Once you've got that you will then need to apply a factor, which for a 980 is approximately 150% of a 680 (this is a rough number as it depends on so many factors), so lets say you need about 0.66 of a 980 for your 1x680. How many do you need for 11 clients?
As for a Titan; Titans (Black) perform slightly worse than a 980 at certain tasks, but they have more vRAM if you think you will need it, so it comes down to how much money you have to throw around and where you want to throw it.

2x8core Xeons?. If you can sell your self on it, cool.

Haha most enterprise storage arrays work off parity so raid 5,6,50 or raid DP yeah I dont use that stuff for my PC but it makes a lot of sense in large environments where raid1 would take too much footprint. I dont use raid now that I'm on ssd. When 15k sas pushes a random workload of 450iops at <10ms and an ssd can do 14k iops at sub ms latency it doesn't make much sense to do in a home PC unless you just mirror.

I'll get some hard numbers to see how scaling out the number of clients impacts vram.

Dual 8 core xeons wasn't an easy sell, but they are $400 more than dual 6 and I want to be sure I dont run into a CPU bottleneck again in the next three years.

Ok so i've got some interesting results from testing.

Once I get up to 5 clients the 2GB vram is full. I continue to stack clients and it seems like the extra load gets pushed off to system memory and system CPU goes up with this as well. So what I did with the remaining 4 clients is set them to use the 770 GPU ... I didn't realize I could set the GPU to one the game is not running on and have the work be done on that GPU. So running this way with clients split between GPUs without isboxer it runs actually really well. I get 50fps and my cpu is also around 65%. Now this is just sitting in my POS. Things could be worse by the time you get on grid with sleepers and start launching drones and shooting. Again without isboxer here is the memory utilization when balanced between cards. I don't think this is possible in isboxer because you need to enable aero and then fps drops dramatically across clients on separate video cards.

GTX 680
2 Clients: 1097MB
3 Clients: 1375MB
4 Clients: 1754MB
5 Clients: 1967MB
6 Clients: 1970MB
7 Clients: 2023MB

GTX 770
4 Clients: 1774MB

Using ISBoxer I've loaded all 11 clients to the login screen. No videofx configuration has been done yet. Video optimized for quality.

2 Clients: 1958MB 62FPS CPU 40% RAM 51%
3 Clients: 2005MB 62FPS CPU 43% RAM 56%
4 Clients: 1966MB 2.33FPS CPU 35% RAM 62%

At 3 clients the system is playable at least without being on grid with any additional ships.... Scratch that. Got into a fight with two guardians and a bunch of battleships. Fortunately I saw the lag start creeping in and disengaged. By the time i was back in my pos I was at 2.33FPS and it didn't get better until I closed an account. So down to 11 clients loaded but only two clients logged in and it was playable again.

At 4 clients the system says nope. The 2.33FPS the monitor shows is probably a generous value. Even web browsing at this point isn't feasible it lags so badly.

This is a lot to think about. Stacked full screen all 11 clients I think the titan black would be necessary due to the 6GB vram capacity.

Another option i guess would be to split the stacks. I know 6-7 runs ok on my 680 so dedicate two monitors on separate GTX980's to stacked full screen ... though this will definitely complicate setup.

How does dxnothing determine which video card to use? I want to make sure the load gets split where one card handles the clients the other handles the dxnothing windows.

Any thoughts or suggestions?

[bob]

Dejavu. http://isboxer.com/wiki/GPU_Management# ... tiple_GPUs

Code:

2. Pick a Window Layout where the Home Region for each Window is on the GPU you would like to power it. That GPU will power the window for its lifetime, regardless of whether the window moves to the other screen via Cross-monitor Swapping.

But of course Cross Monitor Swapping is bad, so don't do it, but videoFX with a source on one and a viewer on the other? No idea what happens there, but I can only imagine that if there is a performance penalty there is would just be a permanent drain.

Other things to note.
Yes CPU and Memory will increase as you load up other game instances. Or are you using perfmon stats to correlate to see if there is some kind of graphics load increase, or context switching penalty?.

Once the vRAM is full, then yes a lot of work will be pushed onto the CPU/system memory, because everything that goes to the GPU has to be loaded and buffered which is all system memory anyway, and the CPU has to spend a lot of time swapping in and out, instead of letting the GPU handle it directly, it now needs management. Just like if your RAM was full the system would spend a lot of time swapping to and from disk instead of real work.

Normally people turn down their graphics quality a bit when multiboxing because it chews GPU cycles and vRAM if you use Ultra/High settings across the board. It's like any tuning, you put in the good stuff on your main and down tune the others. Do you really need to be running MSAA, FSAA, SSAA, TXAA, 16xAniostropic and only knows what else on 11 clients when you aren't looking directly at 10 of them most of the time? (okay you can't run them all at the same time, but they chew vram like it's candy)
Yes, you may switch between preferred slots and so have a couple set higher, but you have to remember the game designers are not designing their graphics usage for multiboxers, it is all with the single player in mind, so Ultra/High graphics settings tend to be fairly intensive even for the one client.


But the good news of course is you have some idea of what is needed for your chosen graphics settings. Although you do need to test in game because the login screens don't load all the game assets and don't have to content with other items that need rendering, like other players, and 50 other ships. You also need to test only with those you are playing , and not have others loaded in the background chewing up background resources that you are trying to evaluate.

Maybe a Titan will be fine, but I suspect if you want to run Ultra/High quality settings on 11 clients even that might not be enough. What you want is someone who owns one and multiboxes EVE to chime in, and not have to read my drivel.

[MiRai]

Using ISBoxer I've loaded all 11 clients to the login screen. No videofx configuration has been done yet. Video optimized for quality.

This is a lot to think about. Stacked full screen all 11 clients I think the titan black would be necessary due to the 6GB vram capacity.

I have to agree with what bob says about lowering your settings, but you also don't tell us whether you're using DX9 or DX11, or even what your GPU load is or what FPS you're running your clients at.

Another option i guess would be to split the stacks. I know 6-7 runs ok on my 680 so dedicate two monitors on separate GTX980's to stacked full screen ... though this will definitely complicate setup.

How does dxnothing determine which video card to use? I want to make sure the load gets split where one card handles the clients the other handles the dxnothing windows.

Any thoughts or suggestions?

It will complicate the setup, because of this:

But of course Cross Monitor Swapping is bad, so don't do it, but videoFX with a source on one and a viewer on the other? No idea what happens there, but I can only imagine that if there is a performance penalty there is would just be a permanent drain.

You're going to need dedicated dxNothing regions per-monitor/GPU or else you're going to take a big hit in performance trying to view VFX Viewers cross-monitor/GPU.