Continuation from http://www.gpugrid.net/forum_thread.php?id=539&nowrap=true#4192

I'm told the behavior in 6.4.3 is correct.

It has been changed because the goal is to have GPU jobs run at top speed without people having to change their #CPUs preference. So if there is a GPU job running, the scheduler will avoid saturating the available CPU's (otherwise the GPU job runs slow).

I hope they changed it only for Windows and not for Linux, otherwise people would end up having one core idle if they use Linux...

I'll give 6.4.4 a try on a Linux box...

It's still not working quite right.

I upgrade to 6.4.4, it runs benchmarks and starts 3 tasks on my quad core with the %processors at 99%. I change it to 100% and 5 tasks start. I close BOINC and restart and it starts 4 tasks while at 100% of cpu's.....

On Windows?

With BOINC 6.4.4 on Vista 64bit and 100% of CPUs I have three CPU tasks and one GPU task running. Like it should work now...

Can you have a look what happens after the next task switch?
____________

pixelicious.at - my little photoblog

XP. On task switch with 100% of CPUs it is running 5 tasks.

Could you post that to the BOINC alpha email list? Maybe they'll listen to you... ;)

I really hope the Synedoche guys will do a better job...
____________

pixelicious.at - my little photoblog

Continuation from http://www.gpugrid.net/forum_thread.php?id=539&nowrap=true#4192

I'm told the behavior in 6.4.3 is correct.

It has been changed because the goal is to have GPU jobs run at top speed without people having to change their #CPUs preference. So if there is a GPU job running, the scheduler will avoid saturating the available CPU's (otherwise the GPU job runs slow).

I hope they changed it only for Windows and not for Linux, otherwise people would end up having one core idle if they use Linux...

I'll give 6.4.4 a try on a Linux box...

They changed it also for the Linux BOINC Clients...
Great! Windows users can now run GPUGRID at full speed because the app gets a full core, but Linux users have now one idle CPU core, because the Linux GPUGRID app doesn't need a full core...

Now Linux users will have to write a cc_config.xml with ncpus+1 again to use all their CPU cores...

That's strange! It seems like they decided this "0.9 CPU 1.0 CUDA" sharing doesn't work and dropped the concept completely? Otherwise the project could just have set 1.0 CPU 1.0 CUDA and they could have achieved the same effect.

With the second method they could implement some way to set different CPU values for different OSes. You could also give the Win / Linux BOINC clients different behaviour, but then everything might get screwed up as soon as there's a second GPU project!

This really has to be set by the project and not by BOINC.

MrS

Yea, that's ok, but there wasn't a difference between low and normal priority with the acemd tasks...
____________

pixelicious.at - my little photoblog

Could you post that to the BOINC alpha email list? Maybe they'll listen to you... ;)

I really hope the Synedoche guys will do a better job...

It's still not working quite right.

I upgrade to 6.4.4, it runs benchmarks and starts 3 tasks on my quad core with the %processors at 99%. I change it to 100% and 5 tasks start. I close BOINC and restart and it starts 4 tasks while at 100% of cpu's.....

It's still not working quite right.

I upgrade to 6.4.4, it runs benchmarks and starts 3 tasks on my quad core with the %processors at 99%. I change it to 100% and 5 tasks start. I close BOINC and restart and it starts 4 tasks while at 100% of cpu's.....

Running 5 tasks could be normal if it determined some task was in deadline trouble.

Hi

I've downloaded "boinc_6.4.4_windows_x86_64". Getting error "Error reading setup initialisation file" when I try to install 6.4.4. I'm running Vista premium 64 bit.
Anybody else with this issue ?

Nope... I also installed it on Vista Premium 64 bit without an error...

Maybe it's only a corrupt installation file because something happened during the download? Have you tried to download the installer again?
____________

pixelicious.at - my little photoblog

Could you post that to the BOINC alpha email list? Maybe they'll listen to you... ;)

I really hope the Synedoche guys will do a better job...

I thought of that.

Can you tell me what cpu (core) percentange you show as being used and the system priority for both windows and linux.

On Windows XP I get about 30-40% CPU (which is 70-80% of one core) at 'Normal'

Could you post that to the BOINC alpha email list? Maybe they'll listen to you... ;)

I really hope the Synedoche guys will do a better job...

I thought of that.

Can you tell me what cpu (core) percentange you show as being used and the system priority for both windows and linux.

On Windows XP I get about 30-40% CPU (which is 70-80% of one core) at 'Normal'

I'll try to do that tomorrow...
But I just read your email you sent to BOINC alpha, you explained the problem pretty good!

I checked in the following change:
- client: tweak CPU scheduling policy. When there's a coproc job:
Windows: don't saturate CPUs
Unix: saturate CPUs

Could you post that to the BOINC alpha email list? Maybe they'll listen to you... ;)

I really hope the Synedoche guys will do a better job...

I thought of that.

Can you tell me what cpu (core) percentange you show as being used and the system priority for both windows and linux.

On Windows XP I get about 30-40% CPU (which is 70-80% of one core) at 'Normal'

I'll try to do that tomorrow...
But I just read your email you sent to BOINC alpha, you explained the problem pretty good!

And the reply is:

"david" wrote:

I checked in the following change:
- client: tweak CPU scheduling policy. When there's a coproc job:
Windows: don't saturate CPUs
Unix: saturate CPUs

of course now you have to wait for 6.4.5, but everybody should be happy then.

Tried download again tonight, now installed & running. Must have been a Gremlin. I'd already tried a second dowload after the first had failed.

I already had trouble with 6.4.4

It ran, one after the other two GPURID tasks, and then they were both ready to report. No attempt to download more work. so the GPU was idle. I had to suspend the cpu only project to force the cleint to start polling projects for work, it then got 1 task for GPUGRID.net and began to run. This is not desired behavior. And it should not even have happened as the two projects running, had equal resource shares and with two cores, that should pretty much keep 1 task for each project running. 6.4.3 and 6.4.2 and before worked better.

I do not know why the decision to change the scheduling policy was made.

Lets hope the future fix applied is a fix.

Today I took delivery of a new graphic card for my second machine. Having installed Boinc 6.4.4 over the weekend all I had to do was to get the new drivers installed and upgraded to the 178.24 level and re-connect the second machine to GPU Grid. One GPU task downloaded and it kicked one existing task out. After a minute or so that task restarted. Since then the GPU task and 4 Cosmology tasks have been happily sharing that machine.

Meanwhile on the first machine the GPU task is sharing with 3 Cosmology tasks.

All the Cosmology tasks on both machines are running high priority (they shouldn't be but Boinc seems to be calculating the work buffer when a GPU task is running the same as if an ordinary project was running).

On the first machine the GPU task is also running high priority because the GPU task will miss its deadline by a few hours.

It looks to me as if the GPU task status is the key to getting 3 or 4 other tasks running, on Windows at least.

Machine 1 : XFX Geforce 8600GTS

Machine 2 : Asus Geforce ENGTX260

Both machines : Vista Home Premium 32 bit, Boinc 6.4.4, Coda 2.0, Nvidia 178.24

Phoneman1

I already had trouble with 6.4.4

[...]

I have trouble with this version and the previous one.

On computers with 2 GPU cards, if this is the same model of GPU card, with the previous version on a quad cores, i had 3 CPU task running and 2 GPU task.
On computer with one GPU card, sometimes i had 4 CPU task and 1 GPU, or 3 CPU task and 1 GPU task.

So with the new version 6.4.3 or 6.4.4, i have on all computers, one task per CPU and GPU. So one quad, 2 CPU task and 2 GPU task, if i have 2 GPU cards.

But, if i have 2 different GPU cards, at one moment, i have to wait that the last task was finished before download another WU. So the more powerful card is waiting.

I downgraded to 6.4.2, because i don't understand why the Task duration correction factor is set to 100 by default.
I understand you want to have full power of the GPU cards, but there is some problems with this configuration.

I can't be monitoring all the time my computers.
I let my computers ran with the 6.4.4 since Sunday morning, and i don't see any improvements with this situation.

Jim PROFIT

6.4.4 works fine. I had to change the Task duration correction to 1.
____________
Constant dripping wears away the stone. :)

6.4.4 works fine. I had to change the Task duration correction to 1.

How do you change the DCF? (XP Home, still with 6.3.21)

The problem with DCF is that when you change it, the value will be change when you upload/download a WU.

So you have to change the value every day!!

You will find this value in the file "client_state.xml".


Jim PROFIT

The problem with DCF is that when you change it, the value will be change when you upload/download a WU.

So you have to change the value every day!!

You will find this value in the file "client_state.xml".


Jim PROFIT

It's more than a bug!
I try the 6.4.5 this night and get the same problem.

This morning i got a computer with 2 GPU, only one GPU was crunching, and the DCF for this computer was set to 100, not by me.
I thought this problem is just on computer with 2 or more GPU cards.

So i downgraded to 6.4.2, because on computer with 2 GPU, the DCF stay all the time to the value of 100. Btu when i downgraded, i have tho manually change this value in the file or the DCF is not change even after one or more WU was calculated!

Jim PROFIT