Multiple GPU tasks einstein@home

[Bradtech519]

Right now I have three workunits running at once on my 580 GTX with CPU tasks disabled. I'm using my CPU for CPU only based projects, and dedicating my GPU to projects that take advantage of GPUs. Are there any kind of recommended settings based on how many GPUs you are running? I've just been watching my GPU usage and it's staying around 95% load, and my temps are between 60-65c. I set my GPU utilization factor to 0.33 in Einstein preferences to get the three tasks. Also, are you running CPU versions of applications for which GPU versions are available?

 

[Sunny129]

Right now I have three workunits running at once on my 580 GTX with CPU tasks disabled. I'm using my CPU for CPU only based projects, and dedicating my GPU to projects that take advantage of GPUs. Are there any kind of recommended settings based on how many GPUs you are running? I've just been watching my GPU usage and it's staying around 95% load, and my temps are between 60-65c. I set my GPU utilization factor to 0.33 in Einstein preferences to get the three tasks. Also, are you running CPU versions of applications for which GPU versions are available?

i contribute to Einstein@Home as well. i crunch 4 BRP tasks simultaneously on a pair of 7970s. i do not run CPU versions for which GPU versions are available, but i do crunch the Einstein@Home CPU-only applications on my CPU (the Gravitational Wave search and the Gamma Ray Pulsar search).

i can't remember exactly what my settings were when i was running 3 GTX 580s, but i do seem to recall that even the 1.5GB version has enough memory for 4 simultaneous tasks...however, you might find that there's no advantage in doing that over just running 3 simultaneous tasks from an efficiency standpoint (you'll have to do the testing and make the calculations to know for sure). your GPU load and temps are spot on with what any one of my former GTX 580s was getting while crunching 3 simultaneous BRP tasks.

if you only have one GTX 580 in that machine (which it looks like you do, according to your computer description at the Einstein@Home site), and you're looking to best maximize both GPU and CPU crunching without letting one hinder the other or vice versa, you'll just have to experiment for yourself. some projects run nicely with others, and some don't...so it really depends on what CPU projects you're planning on running in conjunction with Einstein@Home GPU projects. one of my Einstein@Home 7970s crunches along side another 7970 crunching Milkyway@Home, while the 4C/8T CPU crunches Test4Theory@Home and yoyo@Home, which get just enough resources to not slow the GPU crunching down at all. this allows the CPU to run 1 T4T task and 3 y@H tasks, or 4 y@H tasks, at a time in conjunction with all the GPU activity. my other Einstein@Home 7970 also crunches along side another Milkyway@Home 7970, while its 6-core CPU crunches Test4Theory@Home, Asteroids@Home, and Einstein@Home (both Gravitational Wave and Gamma Ray Pulsar searches). this allows the CPU to run 1 T4T task and 5 A@H tasks, or 6 A@H tasks, at a time in conjunction with all the GPU activity.

note that i can generally run 2 more CPU tasks on one machine than i can on the other despite the fact that they both have 2 7970s doing the exact same things for the exact same projects. of course i'm sure it has alot to do with the CPUs being different (one is a 1090T and the other is a 3770K). also, while all 4 GPUs get full PCIe x16 bandwidth, the 790FX board is only PCIe 2, whereas the Z77 board is PCIe 3.

 

[Sunny129]

i completely forgot about my other machine, which has a single 7970 crunching 4 Einstein@Home BRP tasks at full PCIe 3.0 x16 bandwidth, while the 3770K CPU crunches 1 Test4Theory@Home task and 5 Einstein@Home tasks (both Gravitational Wave and Gamma Ray Pulsar searches), or 6 Einstein@Home tasks, at a time, in conjunction with all the GPU activity.

 

[Bradtech519]

Thanks for the info I appreciate it. I might go up one more task on it. I've been thinking of selling the GTX 580 and getting two 6970s, or one 6990. Might just go up to a new R9 280x which is basically a 7970 though. I missed the boat on Cayman 6900 series cards. There's always the option of pairing up another 580 GTX but it seems the projects that interest me which are astronomy based like ATI/AMD GPUs. The 7970s seem to be doing really well even in Einstein now. Milkyway they dominate in still. I started up Milkyway that I used to run on my 4890 and 5850. Didn't seem to handle it very quick. 5-6 or 12-14 minutes a task. Looked at AMD GPUs and they were doing theirs in 40-50 seconds.

 

[Sunny129]

if you're planning on sticking with Einstein@Home, you're far better off with the GTX 580 you have now than a 6xxx series GPU. a 7970 on the other hand is substantially faster than a GTX 580 at Einstein@Home. i found that out the hard way by investing in 3 GTX 580s for Einstein@Home, not realizing that the 7970 was actually the superior GPU due to a serious lack of test results confirming the 7970's superior performance in that metric. no more than a few months after having purchased the GTX 580s, the project's top hosts were becoming dual and triple 7970 machines...so i cut my losses and took the plunge on some 7970s. i have no idea if the 7970 still rules the roost over at Einstein@Home or if the newer R9 290X/290 is quicker...nor do i know if the Titan or the 780 Ti is faster. no matter though, b/c you can get yourself 3 7970s for the price of a single Titan or 780 Ti and crush them from a PPD/RAC perspective.

 

[lane42]

Question for you guy's, looking at BT's times for his 580, they look to be 10 min.
faster then my 780 ? I'am Running 3 wu at a time@ about an hour, no cpu.
BT's 580 at around 50 min. BT is that for 2 or 3 at a time, or am I seeing things...

 

[Sunny129]

i'm pretty sure BT mentioned in a previous post that he's running them 3 at a time too. but there are several factors that may contribute to the difference:

1) different OSes
2) different driver versions
3) you have 2 780s in that machine while BT only has one 580 in his, and so could be a PCIe bandwidth issue
4) GTX 580 is only marginally slower than GTX 680 on Einstein@Home, and it actually beats out a GTX 670 by a healthy margin, so its quite possible that the nVidia 7xx series GPUs still doesn't do the Kepler architecture justice in the compute department.

...that said, i don't think BT's sample size is big enough to make a judgement at this point - he's only got 17 valid and pending BRP4G tasks combined so far, 11 of which seem to be in the 3100s range, and 6 of which are right in line with your run times of ~3700s.

 

[lane42]

The other gpu in that rig is a gtx 570, it just shows as 2 780's.

 

[Sunny129]

my mistake...regardless, the number of occupied slots is what affects PCIe bandwidth, so its possible that your GPUs are restricted to x8 each (depending on your mobo), while BTs GPU could be getting full x16 bandwidth.

 

[lane42]

yes, 3.0x8 for the 780
2.0x8 for the 570

 

[Bradtech519]

I'm going up to four tasks to see how it goes. I'm letting Asteroids@home run on the eight cores on my FX 8350. It alternates between asteroids and skynet pogs.

 

[Bradtech519]

I just spoke with my master(wife). I might be selling my 580 GTX, and buying the Radeon R9 280x sometime soon. Going to research first, and look around at used 7950/7970 prices.

 

[Bradtech519]

I've hit my milestone of 2M in pogs. I am back to doing einstein@home. I am doing 3 GPU tasks at once on my R9 290. The temp on my R9 290 at 100% load is 61c. I am hoping to not get any invalid workunits. I am not currently receiving any CPU workunits even though it is enabled. I am trying to keep enough CPU resources free for the GPU workunits so I have my use at most setting to 83.3. I wonder why I haven't got any CPU workunits from einstein. Hate to leave cores not at work doing something.

Use CPU
Enforced by version 6.10+ yes Use ATI GPU
Enforced by version 6.10+ yes Use NVIDIA GPU
Enforced by version 6.10+ yes Use INTEL GPU
Enforced by version 7.0.27+ yes
Binary Radio Pulsar Search (Arecibo)
Binary Radio Pulsar Search (Arecibo, GPU)
Binary Radio Pulsar Search (Perseus Arm Survey)
Gravitational Wave search S6Bucket Follow-up #1
Gamma-ray pulsar search #4

Run CPU versions of applications for which GPU versions are available

All checked

 

[Sunny129]

wow old thread resurrection!

i run Arecibo GPU (BRP4G) tasks and/or Perseus Arm Survey (BRP5) tasks 3 at a time on my HD 7970's. i haven't crunched on my single-7970 machine in a while due to video card and PSU issues, so i can't remember exactly how many CPU tasks used to run along side the 3 E@H GPU tasks...but they generally came from the LHC@Home, VirtualLHC@Home, and Yoyo@Home projects. on my dual-7970 machine i run 6 E@H GPU tasks at a time, and i use 4 of my Intel i7 3770K's 8 threads to run one VirtualLHC@Home task and 3 Asteroid@Home or LHC@Home tasks. my CPU is still ~35% idle, but i must have it set that way b/c pushing the CPU toward 100% utilization w/ more CPU tasks causes GPU utilization to suffer.

as for why you're not getting E@H CPU tasks, i'm not exactly sure...your web preferences look like they're set correctly at a glance. try setting the "Run CPU versions of applications for which GPU versions are available" to "No" and see if that helps. this way GPU tasks will run on the GPUs only...there's really no need to run applications on the CPU that were really intended for the GPU. i'm thinking that maybe the way you have it set is confusing the server...

 

[Bradtech519]

Thanks Sunny, they finally came in & started running. I went ahead and unchecked that Run CPU version of Applications for which GPU versions are available. I'm leaving my Laptop on pogs until I hit 2M. I was googling for some questions & actually found my old thread! Thought I'd bump it and talk about getting back into Einstein now. Will be eager to check to make sure my Workunits validate successfully. My CPU load is fluctuating between 65-90% running two GPU tasks & 5 Einstein CPU tasks. I have around a core probably not running but don't want to start getting tons of invalids. I have my Core I5 4200M in my laptop running pogs still, and looking to add another CPU project to a new rig soon. Eventually building up to adding a decent GPU to that new rig next year. The FX 8350 might be relegated to backup rig status depending on Intel's offerings next year & how much the 4790k gets. Those things are killing it on some CPU projects like Universe when comparing run times to FX 8350/9590s

 

[Bradtech519]

Looks like I am getting invalid/inconclusive status on my GPU workunits again. I've lowered my CPU utilization to maximum usage to 75% to see if that helps. I wonder if eventually I will need to go down to 1 Workunit on my GPU at a time. Do you think I should go below 75% on CPU utilization in BOINC?

 

[Sunny129]

if i were you, i would take max CPU utilization as low as 50%. in fact, i might even go so far as suspend all CPU work (from E@H and any other projects that might currently running tasks on your CPU) for the time being just to make sure that the GPU isn't the root of the problem. if you still get invalid GPU tasks while no tasks are running on the CPU, then its probably just a GPU issue. what video drivers are you running right now anyways?

 

[GLeeM]

I wonder if eventually I will need to go down to 1 Workunit on my GPU at a time.

How much of an increase in ppd did you get going from one WU to two WUs?

 

[Bradtech519]

if i were you, i would take max CPU utilization as low as 50%. in fact, i might even go so far as suspend all CPU work (from E@H and any other projects that might currently running tasks on your CPU) for the time being just to make sure that the GPU isn't the root of the problem. if you still get invalid GPU tasks while no tasks are running on the CPU, then its probably just a GPU issue. what video drivers are you running right now anyways?

14.12 Omega Catalyst drivers. I've went through uninstalled the latest boinc & cleaned out the programdata\boinc folder. I went back to 7.2 by accident but I'll just keep using it. I kept getting fed CPU tasks & my machine was doing them regardless of me telling it not to in the einstein preferences. Right now all I told einstein to give me was Aceibo and Perseus GPU workunits. I have two Arecibo's going that will run around 1 hour 43 minutes each & be done. In Task Manager I'm just using 20% CPU. I'll let these run for a while and check the pending workunits to see. Then possibly start engaging one CPU task at a time until I find the systems breaking point. That's if it's not the GPU like you think it could be.

 

[Bradtech519]

How much of an increase in ppd did you get going from one WU to two WUs?

I'm honestly not sure quite yet. I just noticed my GPU was only getting used 30-50% max running one workunit at a time.

 

[Sunny129]

1...That's if it's not the GPU like you think it could be.

when i say that, i want to be clear that i don't necessarily think that your GPU hardware is bad...it could just be your combination of GPU type and driver version. do you need those Omega video drivers for video games?

 

[Bradtech519]

No, just been using the AMD Gaming Evolved program & it recommended updating to them. I do play game off and on but most of them are older games that could run on ancient hardware.

when i say that, i want to be clear that i don't necessarily think that your GPU hardware is bad...it could just be your combination of GPU type and driver version. do you need those Omega video drivers for video games?

 

[Sunny129]

if that's the case, and you don't absolutely need the latest and greatest drivers for the latest and greatest games, then it may be worth it to explore some older driver versions that are known to work w/ Einetsin@Home GPU tasks without issue. i'm not sure what driver versions those might be off the top of my head, but i'm sure you can find recommended stable driver versions if you search the Einstein@Home message boards. for instance, i'm running Catalyst 13.4 on my dual-7970 machine. i would try no CPU tasks first and see if that doesn't eliminate the invalid GPU tasks...

 

[Bradtech519]

Looks like they are still validating as inconclusive. I read on the forums and it seems 13.12 Catalyst driver is good to go with. I'm going to give it a shot and see how it goes.

 

[Sunny129]

yeah, i had a feeling it might be a driver issue...not that anything is conclusive yet until you actually try a different driver version. but at least now you know that the invalids weren't caused by overloading (or even just loading partially) the CPU.