While I'm crunching for GPUGRID my rig freezes. I mean - the whole system (nor BOINC client nor smth else). It makes me a "bit" pissed off, coz I can not work or even browse the internet.

This freezes happens periodically and lasts from couple of seconds to 20-30 seconds. Even if I'm typing on the forums when freeze happened - no letter appear, looks like keyboard hangs also. I do sure this happens when I'm crunching GPUGRID, coz if I'm turning it off - never ever this freezes happen.

From other hand I do not think it's GPUGRID "fault" coz in my understanding BOINC should turn GPUGRID immediately off while I move the mouse or type a single letter,

Of course I tried to turn GPU off using "use GPU while computer is in use" either in BOINC either in web-profile - nothing helps.

Another point, may be it will useful to understand the reason. I'm using "desktop effects" and awn (avant-window-navigator - MacOS style launcher and switches between apps) and in majority cases (but not in all) freezes happen when I'm putting the mouse there in order to switch to smth else. With almost 100% guarantee to freeze the rig is pop-up windows (ever Micky-Mouse size). So, in general ANY activity which requires even small piece of the efforts from GPU (yes, even a letter to appear when I'm typing) causes freeze.

I tried to figure out which app causes this by opening gnome system monitor and then trying to freeze - looks it's Xorg...

Here's my hardware and software:
- E6300 overclocked from 2800 to 4000 (100% stable)
- eVGA GTX275 overclocked to 702/1584/1260 (100% stable)
- ubuntu 9.04 (kernel 2.6.28.15-generic)
- 190.18 and 190.32 driver and CUDA2.3 - was on all of them
- BOINC 6.4.7, 6.6.36, 6.10.4 and 6.10.6 - was on all of them

If you need any additional info or it's necessary to do smth - let me know. Of cource, I'n NOT guru in linux, but not noob :-)

---

OK, forgetting about BOINC and CUDA for a minute...

Are you aware that 190.xx driver series is still considered to be a BETA release on Linux?
Are you aware that other people are reporting 'freezing' problems with 190.xx? There are several threads over on the NV Linux forum.

190.18 intermittent freezes
UseEvents and KDE 4.3 direct rendering broken with 190.16

Let's start here. Do you have 'Option "UseEvents" "true"' in the 'Device' section of your xorg.conf?

---

Crunching on Linux: Fedora 11 x86_64 / nVidia 185.18.36 driver / CUDA 2.2

Sure I know that 190.xx is beta, but I never heard that it's buggy.

I inserted this string into xorg.conf but it becomes even worse: freezes are all the time (or better to say - it's one huge freeze) and sometimes mouse starts to moves, but it's not clicking at all, so I had to reset the rig by button on the case.

There are two things which make me think that the reason behind is BOINC: the 1st - if I'm suspending GPUGRID WU's (but rosetta continue to work) I never ever had any single freeze whatever I'm going (even watching movie). And the 2nd - even if I deselected "use GPU..." elapsed time continue to run and status says "running" but no "waiting to run", so looks BOINC does not intercepting this event and does not suspending GPUGRID...

So there are 2 option: either to wait for good driver (A) or to install 185.xx with CUDA 2.2 (B). Am I right in understanding what u r trying to say? :-) Or should I continue to try to be patient and (when I'm completely pissed of) - just to suspend GPUGRID WU's?

---

There is also a bug with the use while in use options that may mean that BOINC does not get out of the way correctly. You will not see that until post 6.10.9 ...

Not sure if it is applicable or not ... but there is a related bug in BOINC ...

Sure I know that 190.xx is beta, but I never heard that it's buggy.

All Linux nVidia releases are buggy to various degrees! (I don't say that to be argumentative or derogatory, that's just the way it is.) Being an early adopter whn nVidia release new drivers is asking for trouble. Aside from some vdpau fixes or unless you need CUDA 2.3 (which you don't for GPUGRID) and especially as you are wanting to use the machine interactively, if you don't need 190.xx, go back to 185.18.36 and CUDA 2.2.

I inserted this string into xorg.conf but it becomes even worse: freezes are all the time (or better to say - it's one huge freeze) and sometimes mouse starts to moves, but it's not clicking at all, so I had to reset the rig by button on the case.

I didn't actually want you to insert that string into the xorg.conf, just wanted to know if it is there. ;)
If it is there it will exacerbate or actually cause freezes!

There are two things which make me think that the reason behind is BOINC: the 1st - if I'm suspending GPUGRID WU's (but rosetta continue to work) I never ever had any single freeze whatever I'm going (even watching movie). And the 2nd - even if I deselected "use GPU..." elapsed time continue to run and status says "running" but no "waiting to run", so looks BOINC does not intercepting this event and does not suspending GPUGRID...

Right, as Paul points out (in the post above) there is a bug related to 'in use' config options in BOINC, so that's why I said let's leave BOINC and CUDA out of this for the moment.

So there are 2 option: either to wait for good driver (A) or to install 185.xx with CUDA 2.2 (B). Am I right in understanding what u r trying to say? :-) Or should I continue to try to be patient and (when I'm completely pissed of) - just to suspend GPUGRID WU's?

Firstly, I wouldn't recommend running the 190.xx series. If it wasn't for the fact that GPUGRID now requires CUDA 2.2, I'd still be running 180.60 on all my Linux machines. This has been by far the most stable nVidia driver release on Linux for years! (My experience with the first 190.xx driver release was a hard lock within 5 mins. That tells me all I need to know about it. I'll come back again and try again when it is out of BETA and is promoted to the current release.)

Secondly, the video card can only do so much. I do not run desktop effects on any of my crunchers. With a single video card you are asking for and expecting too much. The fancy GUI effects put load on the video card at the same time as it is loaded by the CUDA app. And you expect the desktop to be responsive? My advice, don't do it. I can make my desktop "stutter" with a CUDA app running, desktop effects enabled, and spinning the mouse wheel to scroll in Firefox. Compositing, hardware acceleration, OpenGL, etc. etc. at the same time as CUDA is asking too much.

Thirdly, as Paul pointed out, a fix for the 'in use' option not working correctly. Can you install the latest 6.10.9 BOINC release?

Summary: downgrade to nVidia driver release 185.18.36. Turn off desktop effects. Update to BOINC 6.10.9.

---

Crunching on Linux: Fedora 11 x86_64 / nVidia 185.18.36 driver / CUDA 2.2

There is also a bug with the use while in use options that may mean that BOINC does not get out of the way correctly. You will not see that until post 6.10.9 ...

Not sure if it is applicable or not ... but there is a related bug in BOINC ...

so it's not my stupidity? and there is bug in BOINC. Hope one day it will be fixed...

Paul, where I can get 6.10.9? I try from here:
http://boinc.berkeley.edu/dl/

but there are version for windows and MacOS but not for linux...

Anyway i'm checking this link every day t will be available, i'll post result here immediately.

---

All Linux nVidia releases are buggy to various degrees! (I don't say that to be argumentative or derogatory, that's just the way it is.) Being an early adopter whn nVidia release new drivers is asking for trouble. Aside from some vdpau fixes or unless you need CUDA 2.3 (which you don't for GPUGRID) and especially as you are wanting to use the machine interactively, if you don't need 190.xx, go back to 185.18.36 and CUDA 2.2.

it's pity that all nvidia drivers are buggy, but at least they are not making "normal life" of linux users so terrible like ATI one.

should I remove CUDA completely from the computer and CUDA2.2 on top of 185.xx driver or I do not need CUDA at all? soupid question, I'm really new to CUDA and nvidia...

I didn't actually want you to insert that string into the xorg.conf, just wanted to know if it is there. ;)
If it is there it will exacerbate or actually cause freezes!

oops... At least I tried to fix a how I can replace xorg.conf by previous version from console :)

Right, as Paul points out (in the post above) there is a bug related to 'in use' config options in BOINC, so that's why I said let's leave BOINC and CUDA out of this for the moment.

Hope 6.10.9 version will be available soon and I can check if the bug is still there or not.

[/quote]Firstly, I wouldn't recommend running the 190.xx series. If it wasn't for the fact that GPUGRID now requires CUDA 2.2, I'd still be running 180.60 on all my Linux machines. This has been by far the most stable nVidia driver release on Linux for years! (My experience with the first 190.xx driver release was a hard lock within 5 mins. That tells me all I need to know about it. I'll come back again and try again when it is out of BETA and is promoted to the current release.)

Secondly, the video card can only do so much. I do not run desktop effects on any of my crunchers. With a single video card you are asking for and expecting too much. The fancy GUI effects put load on the video card at the same time as it is loaded by the CUDA app. And you expect the desktop to be responsive? My advice, don't do it. I can make my desktop "stutter" with a CUDA app running, desktop effects enabled, and spinning the mouse wheel to scroll in Firefox. Compositing, hardware acceleration, OpenGL, etc. etc. at the same time as CUDA is asking too much.

Thirdly, as Paul pointed out, a fix for the 'in use' option not working correctly. Can you install the latest 6.10.9 BOINC release?

Summary: downgrade to nVidia driver release 185.18.36. Turn off desktop effects. Update to BOINC 6.10.9.[/quote]

1. I'll install 185.xx version 2morrow (it's bit late now :-) )

2. This is the very last thing I'd lik do understand that the effects causes this freezes, at least for now. I've got ustient (ha-ha) and if I heed to do a lot on my rig I'm just suspending GPUGRID WU's. I know it's not right way to do, but .. this is life.

If due to some reasons this option will not work I'll turn effects off. In fact, I'm almost ready to it.

3. When 6.10.9 will appear somewhere - trust me, I'll be the very 1st guy in the line to get it :-)

and one thing more. thanx a lot for your help to the noob in GPUGRID:-)

---

@CTAPbli here is the address

https://boinc.berkeley.edu/dl/?C=M;O=D

The fix for "in use" bug is NOT in 6.10.9 ... 6.10.9 has minor tweaks for ATI usage only (and these are not usable unless server side changes are made by the project).

6.10.9 is an interim build which is why you don't see a Linux version. For our purposes here there is no difference between 6.10.7 and 6.10.9 (.8 is a bad build).

That is why I said post 6.10.9 ... :)

OK, Paul I can wait :-)

---

The fix for "in use" bug is NOT in 6.10.9 ...

Paul,

I'm confused. Before I open my mouth again, exactly what are you referring to, so I'm on the same page.

The original change to check for a running graphics app....

+// check whether each GPU is running a graphics app (assume yes)
+// return true if there's been a change since last time
+//
+bool COPROC_CUDA::check_running_graphics_app() {
+    int retval, j;
+    bool change = false;
+    for (j=0; j<count; j++) {
+        bool new_val = true;
+        int device, kernel_timeout;
+        retval = (*__cuDeviceGet)(&device, j);
+        if (!retval) {
+            retval = (*__cuDeviceGetAttribute)(&kernel_timeout, CU_DEVICE_ATTRIBUTE_KERNEL_EXEC_TIMEOUT, device);
+            if (!retval && !kernel_timeout) {
+                new_val = false;
+            }
+        }
+        if (new_val != running_graphics_app[j]) {
+            change = true;
+        }
+        running_graphics_app[j] = new_val;
+    }
+}

Or the I: BM 6.10.4 - Cuda task doesn't suspend - The same as in BM 6.10.5 message on boinc_alpha that talks of backing out the check for a running graphics app and reverting to previous behaviour, (if it doesn't work as expected) - which hasn't been done yet?

PS. I had to chuckle this morning. It appears your posts to boinc mailing list are listened to. ;)

+    if (display_driver_version) {
+        sprintf(vers, "%d", display_driver_version);
+    } else {
+        strcpy(vers, "unknown");
+    }

---

Crunching on Linux: Fedora 11 x86_64 / nVidia 185.18.36 driver / CUDA 2.2

should I remove CUDA completely from the computer and CUDA2.2 on top of 185.xx driver or I do not need CUDA at all? soupid question, I'm really new to CUDA and nvidia...

If you use the 185.xx driver you need the CUDA 2.2 toolkit.

oops... At least I tried to fix a how I can replace xorg.conf by previous version from console :)

Did your editor create a backup file, xorg.conf~, that you could move back. Alternatively, just edit it again and remove the "UseEvents" config option. It defaults to "false" if it's not there at all.

Hope 6.10.9 version will be available soon and I can check if the bug is still there or not.

I just asked Paul for clarification in the post above.

1. I'll install 185.xx version 2morrow (it's bit late now :-) )

OK.

2. This is the very last thing I'd lik do understand that the effects causes this freezes, at least for now. I've got ustient (ha-ha) and if I heed to do a lot on my rig I'm just suspending GPUGRID WU's. I know it's not right way to do, but .. this is life.

If due to some reasons this option will not work I'll turn effects off. In fact, I'm almost ready to it.

I understand why you'd rather not turn off desktop events. But the thing is, they tend to expose driver bugs like nothing else can, as well as consuming GPU resources.

3. When 6.10.9 will appear somewhere - trust me, I'll be the very 1st guy in the line to get it :-)

I wasn't aware that a generic 6.10.9 hadn't been built for Linux. In any case, hold off from changing BOINC software versions at the moment. I'd like to understand exactly what Paul is talking about.

---

Crunching on Linux: Fedora 11 x86_64 / nVidia 185.18.36 driver / CUDA 2.2

If you use the 185.xx driver you need the CUDA 2.2 toolkit.

OK. I'll do it this night and I'll post results.

oops... At least I tried to fix a how I can replace xorg.conf by previous version from console :)

Did your editor create a backup file, xorg.conf~, that you could move back. Alternatively, just edit it again and remove the "UseEvents" config option. It defaults to "false" if it's not there at all.

it's pretty easy: I deleted xorg.conf, renamed xorg.conf~ into xorg.conf. Nothing special, really :-)

I understand why you'd rather not turn off desktop events. But the thing is, they tend to expose driver bugs like nothing else can, as well as consuming GPU resources.

So, effects are really helpful for finding different bugs :-) If drivers will not help I'll do it. I'm pretty pissed off.

Anyway, turning this effects off I consider as a temporary, coz in my understanding they should not interfere with BOINC while

I wasn't aware that a generic 6.10.9 hadn't been built for Linux. In any case, hold off from changing BOINC software versions at the moment. I'd like to understand exactly what Paul is talking about.

OK.

---

So, effects are really helpful for finding different bugs :-) If drivers will not help I'll do it. I'm pretty pissed off.

Anyway, turning this effects off I consider as a temporary, coz in my understanding they should not interfere with BOINC while

Desktop effects use hardware acceleration. The same hardware that's running the CUDA code. Your asking the same GPU to compute spinning the desktop cube at the same time as it is crunching numbers with CUDA.

It's great having OpenGL hardware acceleration for desktop effects, offloading the decoding to hardware with vdpau when playing a movie, and being able to crunch with CUDA. But ask a GPU to do all 3 at the same time and it won't do any of them particularly well! I'm not saying don't do it. What I'm trying to say is that if you want to crunch, then dedicate the resource (as much as you can) to crunching. Being able to use a consumer grade graphics card for CUDA crunching as well as displaying your desktop is good to have, but there is a reason that nVidia sell dedicated CUDA hardware solutions that do not have graphics output.

---

Crunching on Linux: Fedora 11 x86_64 / nVidia 185.18.36 driver / CUDA 2.2

First, 6.10.10 for windows is out.

The following two changes are the most relevant and specifically the second one is the one I thought might be addressing the issue raised. Where BOINC does not properly suspend and resume based on preferences. There is another one I saw but even going through the list again manually I seem to be missing it ... sigh ... {edit spelling}

Changeset 19199:

client/scheduler/web: add per-project preferences for whether
to accept CPU, NVIDIA and ATI jobs.
These prefs are shown only where relevant:
e.g., only for processor types for which the project has app versions,
and if it has versions for only one type, no pref is shown.

These prefs affect both client and scheduler.
The client won't ask for work for a device blocked by prefs,
and the scheduler won't send it.

This replaces earlier optional project-specific prefs for
"no CPU jobs" and "no GPU jobs".
(However, these prefs continue to be honored on the server side).

- client: if NVIDIA driver is unknown, say that rather than 0

Changeset 19198:

lient: fix bug in CPU prefs enforcement:
enforce "suspend if no recent input" and "exclusive apps"
only if overall mode if RUN_MODE_AUTO (run according to prefs)

Jack,

As to that code segment and the backing out of the "fix" that is not something I have been referring to anywhere. There *IS* a change in that neck of the woods but it is a change in the logic and involves two flags and is Changeset 19137

Or I have completely lost the whole thread ...

Thanks for the reply.

As to that code segment and the backing out of the "fix" that is not something I have been referring to anywhere.

Well, from the limited testing I've done - check_running_graphics_app(), which is being used for 'Use GPU while computer in use' is not having the desired effect. I need to do some more testing before reporting this.

Or I have completely lost the whole thread ...

No, I don't think so. Too many changes in a short time period.

---

Crunching on Linux: Fedora 11 x86_64 / nVidia 185.18.36 driver / CUDA 2.2

I installed 185.36.128 and CUDA Toolkit 2.2, but freezes still there.

Then I turned off effects. As of now no more freezes, but I still think that it's like hiding the head in the sand.
"As deeper your head in the sand, as more unprotected your ass"
|(c) Ozzi's proverb :-)

Hope Paul will manage to fix this bug one day.

BTW Paul, there is 6.10.10 version for windows only. Hopefully one day linux one will be also available.

---

and for MacOS also, but not for linux. I can wait

---

I installed 185.36.128 and CUDA Toolkit 2.2, but freezes still there.

Then I turned off effects. As of now no more freezes, but I still think that it's like hiding the head in the sand.

Yeah, get a whiz-bang system and then you have to shut off everything to make BOINC work... BOINC is supposed to be working in the background as idle and not interfere with anything ... then again, what do I know ... :)

"As deeper your head in the sand, as more unprotected your ass"
|(c) Ozzi's proverb :-)

I had never heard that one ... or at least I cannot recall hearing it ...

Hope Paul will manage to fix this bug one day.

Sadly guys I can fix nothing... and for the most part my reports are like whistling on the wind ... the good news and bad news is that it looks like my disability is on the upswing again so there will be much rejoycing as I will not be able to do as much ... of course I get the benefit of feeling like I am drunk all the time free of charge ...

BTW Paul, there is 6.10.10 version for windows only. Hopefully one day linux one will be also available.

I mentioned that some place here ... in addition to as I noted above, my attention span is also shot so that is why I am not sure I was tracking this thread well ...

Anyway, based on the notes the 6.10.10 release aside from the UI changes (that may or may not be working) there is nothing in .10 over .7 for us here ... the main development issues in that release is on ATI cards and how they interact with the server. If you want to follow there are notes on Collatz as to what is going on (though they may have referred out to the BOINC board.

In that I am not all that pressed and my set-up is working (with the exception to the FIFO bug, still unacknowledged as far as I know) 6.10.7 would be what I would recommend. The FIFO bug will only get you if you run TWO GPU projects at the same time ... and then the most annoying thing is that it kills the resource share allocations... run time will be biased towards (as near as I can tell) the project that will download the most work for your queue size ...

In my case I am running with 0.1 extra work and it seems to respect shares, sort of ... if MW goes off the air I get several hours of Collatz work instead of only one or two tasks ... I run those off and then I get, if they are back MW work till it goes off the air again ...

I installed 185.36.128 and CUDA Toolkit 2.2, but freezes still there.

Then I turned off effects. As of now no more freezes, but I still think that it's like hiding the head in the sand.

Yeah, get a whiz-bang system and then you have to shut off everything to make BOINC work... BOINC is supposed to be working in the background as idle and not interfere with anything ... then again, what do I know ... :)

Not a lot about CUDA apps? ;)

The CPU component of the GPU task may be idling but the GPU is not idling when a CUDA kernel is executing on it. Desktop effects (compositing) also take resource. (Far more than they should for the sake of eye candy.) Hit the GPU with 'texture from pixmap' while executing CUDA kernels and expect the desktop to stutter.

As an aside, if you read the CUDA release notes, they tell you that individual CUDA kernel launches are limited to a 5 sec run time restriction when a display is attached to the GPU. For this reason it is recommended that CUDA is run on a GPU that is NOT attached to an X display. If you choose to ignore the recommendation, I'd suggest doing everything possible not to add extra load to a GPU while it's running CUDA and connected to a display, like turning off desktop effects.

---

Crunching on Linux: Fedora 11 x86_64 / nVidia 185.18.36 driver / CUDA 2.2