Seti@Home Cuda Sluggishness

I was just wondering if it's "in progress" or at least "under consideration" but I would like the coders to find a way to prioritize the CUDA-accelerated applications to something equivalent to the CPU's "low priority." Is this something possible with the current Nvidia drivers or does it require coding from Nvidia?

My problem is that I just started computing my first CUDA workunits today and my computer (desktop, even as I am typing this) is just laggy and sluggish. I was playing some Team Fortress 2 and I was getting terrible framerates (15fps etc.).

The thing is, I run Boinc on various projects in the background all the time on low priority and my games and computing have so far been unaffected, probably because of the wonderful instruction prioritizing for CPUs that's been built into windows XP. However, there seems to be a problem with prioritizing (or there is no prioritizing?) with CUDA so running CUDA kills performance completely. Sure, it's blazing fast, but I'll have to turn it off when I'm using my computer sadly.

This brings me to an alternate request. If it's too much work, not possible without new Nvidia drivers, or there's another problem, can you please pressure Boinc to include a simple toggle for "use CUDA" in the system tray right-click menu? This way I can turn off CUDA when I use my computer and/or play games and whatnot, while CPU projects continue to flawlessly run on low as they have heretofore done.
____________

I have the same issue in some games...it's more noticeable with slower cards. Only thing you can do at this point is suspend the project that is using CUDA when you want to play a 3d-intensive game.
____________


You will be assimilated...bunghole!

Make a new ticket in Trac. That's the best way to get the attention of the developers, not complaining about it on an (obscure ;-)) Seti forum.
____________
Jord

-BOINC FAQ Service
-Nvidia CUDA & ATI CAL FAQ

Courtesy starts with your first post of the thread.

Make a new ticket in Trac. That's the best way to get the attention of the developers, not complaining about it on an (obscure ;-)) Seti forum.

Now look what you've made us do ;-)

Im having the same Issue since getting the New CUDA work units today my machine has been horribly lagging i had to suspend boinc to be able to do anything
____________

Im having the same Issue since getting the New CUDA work units today my machine has been horribly lagging i had to suspend boinc to be able to do anything

Thanks Richard, that sums it up nicely.

Cheers,

John.

You could open task manager and right click boinc, and the open wu's, change the priority of the tasks and even the affinity of the processors is another idea. Or invest in a threading app :P
____________

so any idea how i can fix this porblem? i've suspended all the Cuda units so i can use my pc everything else works fin just the GPU app... also i noticed the time to copleation instead of going down it seems to go up before i got the new units i was taking an average of 3 min to complete a CUDA unit but had an estamite of 28 min the estimate just keeps goin up it was at 50 min then i suspended boinc
____________

I have a similar thing too.
After exiting BOINC and making sure everything is gone in TaskManager I still cant run FarCry2. It gets to the splash screen and then just flickers with black screens.
Reboot, BOINC not in startup, FarCry2 runs.
Ive also noticed it with Stellarium except it does run but about a 1000th (feels like anyways) of the speed. Same thing .. reboot and alls well.

At the moment I have no CPU based tasks running (6.6.3, no work rx'd) but Ive been peeling the GPU tasks back so fast that there has been no serious difference. The Q6600 is definitely playing second fiddle here and the GPU is only a 8800GTS. The messages tab says boinc estimates it at 41 GFLOPS which compared to my old IBM 4133 days if unbelievable ! :)
Even more incredible is that CUDA has refreshed my enthusiasm for C++

But hey.. its all new stuff and pioneers don't always find the best track.
____________

the problem is this is happining even when im not gaming... just trying to surf the Web is painfull... it's like im trying to run Remote desktop ofer a dialup connection.. everything is painfully slow while CUDA tasks are running
____________

im Using BOINC manager version 6.4.5 and its running Seti@home Enhanced 6.08 ( cuda)

I am no expert at this i've never had an issue with Bonic it's always ran great for me
____________

The performance slowdown with VLARs is known to the application and nVidia developers and it's top of their list to get this one solved, now that the stability is in place. But other than that, there's no news on the next app yet.
____________
Jord

-BOINC FAQ Service
-Nvidia CUDA & ATI CAL FAQ

Courtesy starts with your first post of the thread.

@Byron: It shouldn't matter anymore, no. The scheduler should recognize 1.0 capability as well, but the only way to be sure is to test it.

Edit: ah, so Byron's post was moved or deleted, hence the "wrong thread" message.
____________
Jord

-BOINC FAQ Service
-Nvidia CUDA & ATI CAL FAQ

Courtesy starts with your first post of the thread.

LOL, not to worry. It was just weird to get the error. ;-)
____________
Jord

-BOINC FAQ Service
-Nvidia CUDA & ATI CAL FAQ

Courtesy starts with your first post of the thread.

Try solution from this post:
http://lunatics.kwsn.net/gpu-crunching/ak-v8-cuda-mb-team-work-mod.msg13810.html#msg13810

Well i seem to have solved my CUDA problem. I aborted all the cuda tasks in my list and updated boinc when it downloaded the new cuda units things work fine again

____________

You guys must understand the concept here.......CUDA uses the GPU for a CPU related task so its no longer executing in the CPU, now most newer opererating systems are more graphically based than anything else.

The frequency of the graphics cards and the ram speed of the card and also the chipset speed/ram of your motherboard all are a big factor.

The CUDA driver delivers faster work processing through the manipulation of your pcix pipelines if they are busy even windows is gonna chunk because you have essentially turned your 1000mhz 512mb ddr3 game killing card into no more than a 2meg 100mhz card and given the rest of the power to cpu time!!!

So, unless you can make CUDA disengage which means you'll have to download cuda and non cuda wu's so when cuda is off it can still allow a work flow to happen and also allow the work to be redirected to the spare cpu time.

CUDA Wu's use Folding to manipulate the data to the info is a high load process and unless you water cool ya GPU and overclock them your going to have issues no matter what the only sure fire fix is to make a SLI or a CROSSFIRE type setup with multi gpu' in one machine then from there prioritize Seti cuda Wu's to only use one of the cards somehow and dedicate the other to your gaming.

I'm sure if you had 2 PCIx slots and fitted the 2nd card and didn't plug a screen into it but installed the cuda driver for only one gpu then u would have joy here. (Also make sure you dont run sli or crossfire for this method.)

Sorry bout my spelling/grammar/punctuation etc i was in a big hurry.
____________

See, the seti apps are using 100% of the CPU, but because of the marvel of preemtive SO that 100% has a lower priority that any other app. So when it is time to work, the seti worker makes room for the other apps.
All your consideration about the video card it is not accurate - the GPU has many units, most of them not in use when you are just using a 2D application. So, all the 3D part can process other stuff easy - if you don't play a 3D game. True, doesnt support preemtive multitasking, but you didn't know that something like that is possible.
Anyway, seti app that uses CUDA should be shutdown only when you play a 3D game.
____________

Anyway, seti app that uses CUDA should be shutdown only when you play a 3D game.

In my situation I have some slow-downs, but nothing like I cannot use the browser or my normal PC opperation. I don't have to stop the CUDA for those tasks. And I have running 3 WU at a time - one for each core and one for the GPU.
Just in games (like Crysis) I feel the pain :)
____________

Sometimes throughout the workunits it seems like there are "bursts" of responsiveness i.e. sometimes the system starts drawing windows faster, but most of the time there's annoying lag. Even as I'm typing this, there is lag between the keypresses and when the letters appear.

Currently, I have to go in Boinc and manually suspend all Cuda workunits every time I get on my PC. Needless to say, it is very rapidly getting on my nerves...

I too find most of the slugishness to be at the start or end of a typical workunit. However, VLARs are choppy throughout the entire workunit. I'm told that will be fixed in an upcoming version, so for now I simply suspend them and resume them before I got to bed so they can crunch without bothering me.
____________


You will be assimilated...bunghole!

Double post...

Hm...

Weird. I "downgraded" from drivers 181.22, because I CANNOT play left 4 dead with them. The game was stuttering heavily and crashing (after random periods, sometimes 10 minutes sometimes 2 hours) before I disabled "multicore rendering."

Disabling that solved the stuttering, but the crashes remained. They were some pretty weird crashes too - the monitor would suddenly go to a solid (primary/screen average at time of the crash) color and the last 1/4 seconds of sound would loop. The hard drive would grind for 5 seconds and then stop. The sound would continue looping. Only way out, hard reset. No response from ctrl+alt+del and then alt+u+r...

Then, the weirdness - my system would reset continuously after soft-reset. It wouldn't even go to the bios screen just reset, reset, reset every second. The screen - flashing... only way out - hard turn off and then turn on with holding the power button.

I'd be inclined to think it's a lack of power problem, but my PSU is 620W and high-quality (Corsair HX620), and I'm drawing less than 300W at both CPU and GPU 100% loaded at the wall per my lame socket-plug watt meter... AND also I've been running left 4 dead 100% A-OK with 177.92 (my previous "preferred" set which also doesn't mess up AA on Battlefield 2142... lol).

So anyway, after downgrading to 177.92, I had some lesser sluggishness this morning, and now it's perfectly fine.

Then again, this might be due to different types of workunits so IDK.

My computer is definitely not slow... Opteron 170 at 2.7Ghz. The ram (2GB) is also running 1:1 with the processor FSB so quite fast. The GTX 260 Core 216 is running at 679/1464/1080 at 24/7, both are 100% stable under continuous 100% load... so it's probably either the drivers, workunits or a combination of both.

I'm going to continue to suspend for games nevertheless :-/

I hope nvidia sorts out their problems with drivers 180 and left 4 dead.

Keep up the good work on Seti@home and Boinc ;-)
____________

Hm...

Weird. I "downgraded" from drivers 181.22, because I CANNOT play left 4 dead with them. The game was stuttering heavily and crashing (after random periods, sometimes 10 minutes sometimes 2 hours) before I disabled "multicore rendering."

I did not see where this question was asked...... Were you running seti CUDA at the same time? You should not run graphics intensive apps and seti CUDA at the same time.

Im running windows 7 and cuda im happy to report im getting no sluggishness what so ever anymore!!!

Trying to get some data for you all to look at and find out why windows 7 fixed this issue.
____________

...and here it is under the name "Suspend GPU work while computer is in use? Enforced by version 6.7+"

Or you could always give Windows 7 a go!!!! Im sure you will be shocked at the performance difference.
____________

In this and other threads, a common thread (pardon the pun) is that people would like to be able to control when BOINC may use the GPU separately from when BOINC may use the CPU.

The current features for v6.6.20 are:

1) Suspend use of the GPU when the computer is in use. This is settable in the UI Preferences.

2) Suspend use of CPU and GPU when the computer is in use. Settable in the UI Preferences.

3) Suspend use of CPU and GPU, selectable from the System Tray.

4) Suspend use of CPU and GPU while some other program (ie a game) is running. Settable in cc_config.

Each of these features does a part of the job, but none of them can do it all, and no combination of these features can do it either.

The problems are:

1) This means that if you move the mouse, or touch the keyboard, the GPU process is suspended for some user specified period of time (usually 1 min). However, this is way too much sensitivity to user input. Moving the mouse to wake up the screen and see if you have received any e-mail is not such a CPU or GPU intensive operation that the BOINC processes should be stopped. Also, in this mode, I have seen the GPU process wake up in CPU Fall-Back mode (100% CPU usage, 0% GPU usage) after being suspended. So constantly suspending GPU processing for every mouse movement is counter productive for BOINC, and for any other process that needs CPU time when the Cuda app starts using 100% of CPU time at normal priority.

2) All that is wrong with 1) goes double here!

3) This is overkill. CPU multi-tasking is quite able to handle BOINC and other processes in all but a very few cases, without resorting to suspending the BOINC processes.

4) All that is wrong with 3) applies here too. When you have a multi-core CPU, and a GPU, and you want to play some 3D-intensive game, then the Cuda processes may need to be suspended, but the CPU processes can still run on the unused CPU(s) and even in background on the same CPU as the game.

5) And finally, there needs to be some consideration of multi-GPU systems on which it is desireable to select which GPU(s) BOINC is allowed to use, and when their use should be suspended. Put a 900+GFlop NVidia Tesla into your system alongside your 300+ GFlop GeForce 8800 GX, then watch options 1) and 2) shut down both the GeForce and the Tesla when you move your mouse! 8^O

To quote the Trac Ticket #842 ...

Ticket #842 (closed Enhancement: wontfix)

Two controls are suggested:

1) User interface in system tray, similar to current 'suspend' or 'snooze' (but acting on GPU processes only)

2) "exclusive_GPU_app" tag for cc_config.xml, analogous to existing "exclusive_app" (implies exclusive use of CPU) - allows automatic suspend/resume of GPU tasks while specified user applications are running.

• There's no place like 127.0.0.1
  
• The best way to accelerate a Mac is at 9.8 m / sec^2
  
• Firefox Personas