Regarding NATHAN WUs:
I believe the issue some of you are having is caused by high memory usage of Nathan WUs running on cards with less RAM. This is a hypothesis, but I have noticed the following:

* TONI WUs (long) e.g. WU 3928763 causes 92% GPU load, and approx 500 MB GPU memory use.
* NATHAN WUs (long) e.g. WU 6183375 causes 96-98% GPU load, and approx 1100 MB GPU memory use.

I am having no trouble at all running NATHAN WUs, but all of my cards have +2GB of memory. I can imagine that if your card has less than arround 1100 MB of RAM, you would face some problems like the ones you have with your 768MB GTX460. When I read the Recommended GPU FAQ I notice that the GTX460 is only recommended for use with short tasks, and that all GPU in the category "highly recommended" has 1GB of RAM or more.

I found the GTX460 768MB was fine with all long WU's except for NATHANS, which needed more memory to run.

Normally GPUGrid tasks scale down well; require less GDDR on smaller GPU's.
I don't have a GTX460, so I cannot check how much memory these tasks use.

---

FAQ's

HOW TO:
- Opt out of Beta Tests
- Ask for Help

Sorry for the late reply. I did not know about this thread until a few days ago, and I just got around to replying now. The original subject matter wasn't directed at me, and so I wasn't following it.

As Bjarke and TheFiend have said, the likely issue is that the 460 does not have enough memory. My simulations are typically of larger biological systems than those of the other scientists, meaning more on-card memory is required for them. If the entire simulation can't fit on the card memory, a secondary algorithm must be used which is slower. There is not much we can do about this.

However, I don't think that it results in a 50% slowdown (1/2 speed). Tomba, for your case specifically, I have pulled some stats from the database (for the period from Nov6 to Nov19, the period I could query right now):

Total: 23 WUs returned, 12.7 hours average runtime, average 6605 credits/hour
Non-NATHAN tasks: 13 WUs returned, 11.7 hours average runtime, average 7186 credits/hour
NATHAN tasks: 10, 13.9 hours average runtime, average 5971 credits/hour

So, it looks like it causes a ~17% slowdown, not 50%. Obviously not insignificant, but there isn't anything I can do about it. A few percentage of that might be due to other factors as well. Since it only affect a small percentage of our users (those with low memory cards), and only causes partial slowdowns for them, we choose to proceed. The BOINC server doesn't allow a huge amount of leeway with regard to what machines get which WUs, so our only choice is to continue or not do the research at all.

Again, sorry for the delayed response.

Hello Nathan! Welcome to our thread (at last!!!).

Sorry for the late reply. I did not know about this thread until a few days ago, and I just got around to replying now. The original subject matter wasn't directed at me, and so I wasn't following it.

There is a moderator posting to this thread. I would have thought that he/she would have notified the concerned scientist (you) in less than the five months this thread has been active.

As Bjarke and TheFiend have said, the likely issue is that the 460 does not have enough memory. My simulations are typically of larger biological systems than those of the other scientists, meaning more on-card memory is required for them. If the entire simulation can't fit on the card memory, a secondary algorithm must be used which is slower. There is not much we can do about this.

Slower... (?)

However, I don't think that it results in a 50% slowdown (1/2 speed).

That is not the issue here. I absolutely don't have a problem if my modest GPU hardware causes a slowdown of your WUs. What I and many other volunteers are experiencing is a lock-out of our PCs. Normal Windows operations take 10-times longer with some of your WUs.

Tomba,

That's me!

for your case specifically, I have pulled some stats from the database (for the period from Nov6 to Nov19, the period I could query right now):

Total: 23 WUs returned, 12.7 hours average runtime, average 6605 credits/hour
Non-NATHAN tasks: 13 WUs returned, 11.7 hours average runtime, average 7186 credits/hour
NATHAN tasks: 10, 13.9 hours average runtime, average 5971 credits/hour

That's all very well, but did you also notice that, in that time period, I aborted no less than 30 of your WUs because they were killing my PC performance?

Dear Nathan. Pretty please, have another look. Thank you, Tom

---

I noticed that some NATHAN WU's result in stopping the CUDA program and thus error out. Mostly that are long runs.
However a lot of these NATHAN's run complete.
And I have also had TONI's that result in error by stopping (or killing) the CUDA program.
But I guess this happens once in a while as I see it also with other projects. It's all in the game, and this game is important.

I have a GTX285 and a GTX550Ti by the way.

---

Greetings from TJ

I can't tell exactly what's going on with your machine since I'm not sitting there using it, but from what I can tell reading this thread the likely problems are:

(1) You have one GPU, which is naturally your display card as well. You are attempting to crunch GPU tasks and use the computer at the same time. This is generally not a good idea with an older card where your display requirements are heavy, especially when watching movies or doing 3D work, and (2) if there isn't enough memory on the card for the WU, then the program/OS will look to your system RAM, and then to the swap memory (Virtual memory on Windows). If it is doing the latter, performance will be catastrophically bad, because it is using the hard drive as RAM. It sounds like this might be what is happening if it is "killing" your machine.

To resolve the first problem, you should consider suspending GPU tasks while the computer is in use. To resolve the second, when the machine is not in use make sure that any programs that use significant memory, and don't need to be running, have been closed. This would hopefully keep large, efficient (98%+ gpu usage) work units such as mine from causing systems like yours problems. My final suggestion is for you to crunch only on the short queue. I do not have work units there, but even so I can't promise there are no large, memory heavy, efficient work units there either.

Please don't assume that I don't believe that you are having serious problems, or that it is not an issue on our side. However, if the problem goes beyond what is mentioned here, then it is truly a problem that can not be fixed anytime soon. For example, a memory leak caused by the application or the Nvidia driver would be an example. I have mentioned the possibility to the programmers here, but they are busy and an update would take time. We can only update the application once or twice a year for safety and continuity reasons. And obviously if it is an obscure driver error, we have no control over it.

If you have any additional information, such as what happens to the memory when you see this problem, then we might be able to help more, or at least resolve it in future versions of the application.

Hello Nathan. Thank you for your insights.

(1) You have one GPU, which is naturally your display card as well. You are attempting to crunch GPU tasks and use the computer at the same time. This is generally not a good idea with an older card where your display requirements are heavy, especially when watching movies or doing 3D work

I don't do movies or 3D work.

(2) if there isn't enough memory on the card for the WU, then the program/OS will look to your system RAM, and then to the swap memory (Virtual memory on Windows).

I have one gig on the card, and six gigs of RAM driving an i7. Right now I'm running I6R9-NATHAN_RPS1120913_respawn2-65-100-RND8231. I don't have a performance problem. Here's what my gadgets tell me:



I guess I'll wait for a killer WU and report again.

FYI, in the past two months I have had one non-Nathan WU that killed my PC: 2CN4_3-GIANNI_bias-11-25-RND1144

---

This thread started off as one of several threads about the 3.1 vs 4.2 app. It then went to SLI and while not quite hi-jacked, as the initial problem was solved (and the title fits), it’s moved onto what seems to be a general complaints thread, about performance issues and limitations of the GTX460, amongst other topics.

I think it’s important that the researchers specify the requirements for various task types, so people can opt in or out of the project, or for long or short tasks.
Of late we have had some heavy GPU memory requirements for several lines of research, some WU’s requiring too much CPU to facilitate even a mid-range GPU and untested release failures. Not much point adding Twitter if you can’t Tweet the basics in the forum!

I wasn't aware that there are presently work units requiring >800MB for normal system performance. These things are normally identified by testers, following testing. Mind you, nobody posted how much GPU memory their GPU’s were actually using, and you can’t see this in the standard error output.
Anyway, these are difficult situations to resolve without having some sort of task-selection mechanism.

If this is a bigger problem than GTX460’s only, I would suggest such heavy-memory or CPU overly-dependent tasks be run only from the Beta queue, or some other test queue for wanton/extreme tasks. While I commend expanding the boundaries of research, I would prefer more choice.

If the server excluded CC2.1 cards, then it would exclude all those GTX460's with 1GB GDDR, and other CC2.1 cards with large amounts of GPU memory. The situation is frustrating, and that wouldn't make it any less so. Perhaps CC2.1 for short tasks only?

tomba, W7 allocates 10% of the GPU’s memory to itself and allocates this ‘dynamically’. Historically, this Desktop Window Management system, results in an 11% or worse performance loss for GPUGrid tasks on W7/Vista, and that’s after disabling Aero and making system optimizations. Note 2008servers are around 4% and 2003 servers operate like an already optimized version of XP.

On my GTX470 a NATHAN_RPS* (non respawn) task uses 960MB GDDR5 on a 2008R2 server, as did the next such task. On W7 it required 1010MB. The GTX470 has 1280MB (roughly), so the 10% hit doesn't effect these WU's on a GTX470 on W7. However, 1GB less 10% would put your GTX460 1GB into the category of 'not enough GPU memory' (921MB), basically because of the Operating System. This would not be the case on XP or Linux and it's less on a server.
There are other considerations that could be at play, such as occasional downclocking, not having enough CPU free to support the GPU, having limited system memory, or running heavy disk I/O tasks, but unless you specify these details I can't make any such determination. The same applies for other moderators/testers and researchers.

From what you report, the respawn tasks use substantially less GPU memory, but there are at least two types of non-respawn NATHAN tasks. I don’t know if both, or just one has the memory requirement problem?

---

FAQ's

HOW TO:
- Opt out of Beta Tests
- Ask for Help

Perhaps CC2.1 for short tasks only?

I do NOT agree with you. I have got one of the first GTX560Ti CC2.1 1023MB and I have no problem on 64bit XP crunching any longrun task. Some teammates have GPU CC2.1 with memory over 1,2GB and they did not report any problem in our forum or here. By my point of view there is no reason to disable longrun tasks for CC2.1 due to problem of one or several specific systems.
In all I agree with you, especially in point of necessity to specify system having problems exactly.
Off Topic for tomba: How do you like WCG HCC GPU tasks on nVidia GPU?

nenym, You are quite right, if the issue isn't just with 768MB cards (or less), then it could be with numerous 1GB cards (on Vista/W7) when running these particular NATHAN tasks.
While there probably isn't any problem with 1GB cards on Linux, XP, 2003 server, and possibly not on 2008R2 servers (touch and go with that one), it's almost certainly not isolated to GTX460 768 cards. It would certainly impact on the 512MB GT440 and GTS450 and probably most 1GB cards on W7/Vista including the GeForce GTX465 (1GB), four types of GTX460, two types of GTS450, some GT440's, GTX560Ti, GTX560, GTX560SE, GTX555, GTX550 Ti, GTX545, GT530, and most disturbingly it could cause problems with some GF600 cards; GT640, GTX650, and even the GTX650Ti.

Thanks,

---

FAQ's

HOW TO:
- Opt out of Beta Tests
- Ask for Help

In that case I see one way you have mentioned yet - fourth queue for the video memory consumption longrun tasks. One could choose if run these tasks or not.

Off Topic for tomba: How do you like WCG HCC GPU tasks on nVidia GPU?

Me no understand "WCG HCC"!!

---

Off Topic for tomba: How do you like WCG HCC GPU tasks on nVidia GPU?

Me no understand "WCG HCC"!!

World Community Grid, project Help Conquer Cancer.

Off Topic for tomba: How do you like WCG HCC GPU tasks on nVidia GPU?

Me no understand "WCG HCC"!!

Then watch this.

Hi everybody !

The title of this thread is "Very bad display performances" and was opened by me. It would be fine if I could add "with SLI and/or multi-GPU mode enabled"
For me this issue is SOLVED : If , when GPUGRID is running , your display refresh is VERY slow as if the system was slowed down, first uncheck SLI and/or multi-gpu performance modes if, of course, applicable to you.
If this is not sufficient or applicable, please open a NEW thread.

WU programming vs GPU capabilities is off topic.

Regards.

---

Lubuntu 16.04.1 LTS x64

I found what looks like a comprehensive GPU watcher; GPU Monitor:



Here we have two NATHANS with a TONI in the middle. None of then gave me PC lock-outs.

Left clock is CPU - little running in my i7 - and the right one shows less than 50% usage of my six gigs of RAM.

The left pair show almost identical GPU status. The other is using almost all my 1 gig of GPU RAM; only 13MB free. But it still did not interfere with normal PC processing.

I'm almost looking forward to killer NATHAN!!

---

Oops - duplicate...

---

I'm almost looking forward to getting a killer NATHAN!!

I got one:



No effect on CPU or PC RAM, nor on most GPU measures. The only difference from the other three is that the GPU load is 99%. That seems to be what's killing my PC performance. It's nothing to do with GPU RAM (357MB free).

---

As Bjarke and TheFiend have said, the likely issue is that the 460 does not have enough memory. My simulations are typically of larger biological systems than those of the other scientists, meaning more on-card memory is required for them. If the entire simulation can't fit on the card memory, a secondary algorithm must be used which is slower. There is not much we can do about this.

However, I don't think that it results in a 50% slowdown (1/2 speed).

In my tests it's almost exactly a 50% slowdown on the GTX 460/786MB cards. It's a memory size problem since the GTX 460/1024MB cards do not have the 50% performance hit. ON the longer NATHAN WUs there's no way to make the 24hr deadline with cards with less than 1024MB either. The TONI WUs run in 12.5-13.5 hours on my 4 GTX 460/786MB cards. As suggested, a solution would be a separate queue for WUs requiring a large memory footprint. Should be simple to implement.

Well I have a GTX285 with 1Gb of memory, Vista x64 ultimate and driver 285.62.
This pc does short runs and long runs from Toni and Nathan and finish them between 65.000 and 78.000 seconds. Perhaps a slow down but finish with good results.

I had updated to latest driver and that resulted in a lot of errors so I downgraded again.
Now I have sometimes a problem that BOINC seems to hang and no respons of mouse of keyboard is possible.
When suspending GPUGRID, same happened and was Rosetta crunching. Booting and Rosetta suspending, GPUGRID running, also freeze after a few minutes.

Booted again, cleared my register, did a full virus scan and removed a piece of software (3.3 was the name) what I didn't know what it was or did. But afterwards the systems in crunching Rosetta and GPUGRID for already 23 hours and 8 minutes without issues or freezes.

For Tomba: perhaps you can check with process explorer (from Microsoft, free to download) if other applications are occupying your system.

---

Greetings from TJ