Hi!!

I upgraded to boinc 6.13.12 and now OpenCl on my ATI card works

I use Catalyst driver 11.3 and SDK2.4 and they work "perfect" with all other ATI boinc projects.

I find the usage of the GPU extremely low. And GPU are only running for a fraction of a second every minute....

CPU usage is also very low.

Is there any way that you can increase GPU load?

Any chance that an app_info.xml example can be posted here on the forum?

---

I use Catalyst driver 11.3 and SDK2.4 and they work "perfect" with all other ATI boinc projects.

According to ATI themselves, SDK 2.4 requires Catalysts 11.4

I find the usage of the GPU extremely low. And GPU are only running for a fraction of a second every minute....

CPU usage is also very low.

Is there any way that you can increase GPU load?

Just wait until the task gets on its way in earnest. I've seen loads of 43-50% for the first 15 minutes, with 96% after that.

Any chance that an app_info.xml example can be posted here on the forum?

Why? What do you expect of an anonymous platform file to do that we're not already testing under normal circumstances? Why is it that you're here?

---

Jord.

BOINC FAQ Service

They say most of your brain shuts down in cryo-sleep. All but the primitive side, the animal side. No wonder I'm still awake.

Reply to ageless


I've been using Catalyst driver 11.3 with SDK2.4 for a long time, maybe 6mounths on 2 different machines with 2 different ATI cards.
With thoose machines I've been running without any problems, Milkyway, Collatz, Primegrid(pps)ati-application, SETI(lunatics ATIbeta application), SETIbeta(Raistmer beta applications(several of them).
And the performance of the tasks in all of theese projects are Very Good!!!

Whatever ATI says, I recommend running Catalyst 11.3 with SDK2.4
Maybe one need 11.4 and SDK 2.4 for some special games that their cards support...


I will run a couple of tasks and I will monitor how they use my GPU

I've seen an app_info.xml been used in several different purposes.
You can use it to adjust different settings on different rigs to improve performance.

On the question why I am here:

I saw a discussion on a different forum that ALBERT where testing OpenCl on ATI and since I have 2 Computors with ATI i'd joined....

---

Maybe one need 11.4 and SDK 2.4 for some special games that their cards support...

Why not read the link? It's the requirements to run OpenCL on the ATI GPU. Of the projects you name, only Primegrid uses an OpenCL application, and then only for their Proth Prime Search (Sieve). You're running the CUDA version, thus nVidia. Not OpenCL. Neither Collatz, nor Milkyway, nor Moo! Wrapper run OpenCL applications. They run CUDA and CAL only.

The other projects where ATI applications are used are:
- Seti Beta: default application for Astropulse is CAL/Brook+ (recognizable by "ati13ati" or "ati14ati"), not OpenCL.
- Seti anonymous platform for Astropulse and Multibeam. However, for the Astropulse application it depends on what you chose to run, for they've got the hybrid application (runs CAL/brook+ and CPU) and an OpenCL application. Mulitbeam are all CAL/Brook+, not OpenCL.

So for OpenCL it does matter which drivers you run in combination with which SDK. For Cal/Brook+ it doesn't so much, as long as the drivers are older than the 8.12

I've seen an app_info.xml been used in several different purposes.
You can use it to adjust different settings on different rigs to improve performance.

For CUDA & CAL, maybe yes, depends highly on what client version you use and whether or not the project runs with the CreditNew scheme or not. For OpenCL, not so much as the project has to send along all the details about the task you run. That's what we're testing here, not our own values.

We're testing OpenCL applications here, not set some app_info.xml and forget all about it. This project isn't for credit hunger people, they say so themselves on their front page. It's very well possible that a new application is released every day; that's undoable with an anonymous platform file.

---

Jord.

BOINC FAQ Service

They say most of your brain shuts down in cryo-sleep. All but the primitive side, the animal side. No wonder I'm still awake.

Maybe one need 11.4 and SDK 2.4 for some special games that their cards support...

Why not read the link? It's the requirements to run OpenCL on the ATI GPU. Of the projects you name, only Primegrid uses an OpenCL application, and then only for their Proth Prime Search (Sieve). You're running the CUDA version, thus nVidia. Not OpenCL. Neither Collatz, nor Milkyway, nor Moo! Wrapper run OpenCL applications. They run CUDA and CAL only.

The other projects where ATI applications are used are:
- Seti Beta: default application for Astropulse is CAL/Brook+ (recognizable by "ati13ati" or "ati14ati"), not OpenCL.
- Seti anonymous platform for Astropulse and Multibeam. However, for the Astropulse application it depends on what you chose to run, for they've got the hybrid application (runs CAL/brook+ and CPU) and an OpenCL application. Mulitbeam are all CAL/Brook+, not OpenCL.

So for OpenCL it does matter which drivers you run in combination with which SDK. For Cal/Brook+ it doesn't so much, as long as the drivers are older than the 8.12

I've seen an app_info.xml been used in several different purposes.
You can use it to adjust different settings on different rigs to improve performance.

For CUDA & CAL, maybe yes, depends highly on what client version you use and whether or not the project runs with the CreditNew scheme or not. For OpenCL, not so much as the project has to send along all the details about the task you run. That's what we're testing here, not our own values.

We're testing OpenCL applications here, not set some app_info.xml and forget all about it. This project isn't for credit hunger people, they say so themselves on their front page. It's very well possible that a new application is released every day; that's undoable with an anonymous platform file.

OK

Then I might getting some feedback of which driver in Combination to which SDK that works best for this project.

ATI Cards I have access too is:

5850 Driver=? SDK=?
4650 Driver=? SDK=?
4650(agp)Driver=? SDK=?
2600 Driver=? SDK=?

The 2600 isn't that important....Doesn't run much on it...
Same with the 4650(AGP).
And someone that runs 68xx and 69xx might want to know what combination will work best for them.

Whatever AMD/ATI says and what you say with your expertise in programming.
I as a user say that so far until proven otherwise I stay with the 11.3 and the SDK2.4 that so far has worked perfectly with the projects you have discribed above wheather it has ati13 ati14 CAL/BROOK+ OpenCl.
They all have work with very good performance compared to other combinations of the driver and SDK that I have seen.

I find it also useful when I doesn't have to change driver/sdk when running different projects.

---

I forgot to mention that today I installed the 11.10 driver with SDK2.5
And I couldn't see any difference from using the 11.3 with sdk2.4

I just wanted to mention that....

---

I also read the ATI page you submitted...

I quote them "The SDK may work on configurations that fall outside the list below."

Well, then I guess I am just lucky getting better performance with 11.3 and sdk2.4 for unknown reasons, well atleast unkown for me.

I run it on 3 Computors and I don't see any problems when doing so

By putting them together you make them look like they're all one and the same thing. They aren't.

ati13 ati14 CAL/BROOK+ OpenCl

ati13ati = ATI 1.3, is the version number against which ATI CAL API the science application was built. In this case ATI CAL one point three.
ati14ati = ATI 1.4, similarly as to above, but then for one point four.
These version numbers coincide with the minimum driver version one needs to run them at. 1.2 was Catalysts 8.12 and above, 1.3 is Catalysts 9.2 and above, and 1.4 is Catalysts 9.3 and above.

For a nice night of reading,
CAL/Brook+/BrookGPU = Compute Abstraction Layer, http://en.wikipedia.org/wiki/BrookGPU.

CUDA = Compute Unified Device Architecture, http://en.wikipedia.org/wiki/CUDA.

OpenCL = Open Computing Language, http://en.wikipedia.org/wiki/OpenCL.

And finally, do read this very interesting article, OpenCL vs CUDA misconceptions.

---

Jord.

BOINC FAQ Service

They say most of your brain shuts down in cryo-sleep. All but the primitive side, the animal side. No wonder I'm still awake.

That wasn't helpful information at all...

All I wanted to know was if the behaviour of the application i discribed was due to driver or sdk malfunction in some way...

And or if it is something that I can do somehow to change that behaviour.

A simple "NO you can't" from a program developer would have been a nice reply.

You seen to have great knowledge in where to find information to read.(Librarian).

Why you(ageless) are here I haven't figure out yet.

---

I am not a program developer, but just another volunteer with perhaps a little more information in his pocket than others have. I try to teach. You don't seem interested to learn, so I won't teach you anymore. I'll now ignore you. Your loss.

---

Jord.

BOINC FAQ Service

They say most of your brain shuts down in cryo-sleep. All but the primitive side, the animal side. No wonder I'm still awake.

I can't loose anything if I don't have gained anything....

What did I loose....?

Hmm.

Ok, Goodbye then.

---

Is there any way that you can increase GPU load?

No, not with current Apps. Actually so far we are quite happy we got it working at all.

Any chance that an app_info.xml example can be posted here on the forum?

For now I wouldn't recommend to run Apps on anonymous platform here. We may need to release and deprecate apps at a pace that manual updates won't be able to keep up with.

We may ask you to try some manual tuning later, but only after the development has settled a bit.

BM

Thank you for that answer.

Will you have a thread in the forum about application updates when they arrive and some information of what can be expected from it?

---

Is there any way that you can increase GPU load?

I'll try to answer this question.

Yesterday I ran the app 1.19 (atiOpenCL). My computer is AMD Phenom X4 925. Two ATI HD4850 1GB. Software is Windows 7 Ult x64. Boinc 6.13.12. AMD Driver 11.11
Screenshot

It seemed to me that the GPU load is very low. I looked at the Gigabyte HUD program and saw that the load on the GPU is ~ 40%. And progress of WU increased by 0,002% in the second.

I tried the following method. Priority of the process - real time.
Screenshot 1
GPU load was ~ 80%. And progress of WU increased by 0,003%-0.004$ in the second.

Then I returned the priority - below average. GPU load returned to ~ 40%. Progress of WU increased by 0,002% in the second.
Screenshot 2

Then I again tried real time priority.The result was repeated. GPU load ~ 80%.
And progress of WU increased by 0,003%-0.004$ in the second.
Screenshot 3

The screenshots you have to look at Image

Hi.

I see you have a runtime of aprox. 10hours on your 4850.
I use a 5850 and have a runtime of 10hours as well.

Did the runtime decrease with 40-50% when you ran the task in "real-time mode"?

I tried to look at your computer but it hadn't any completed task so i couldn't look at it. I'll get back to your computor in a day or 2.

http://albert.phys.uwm.edu/results.php?hostid=1393

Interesting point there with increased priority.

//TQ

Is there any way that you can increase GPU load?

I'll try to answer this question.

Yesterday I ran the app 1.19 (atiOpenCL). My computer is AMD Phenom X4 925. Two ATI HD4850 1GB. Software is Windows 7 Ult x64. Boinc 6.13.12. AMD Driver 11.11
Screenshot

It seemed to me that the GPU load is very low. I looked at the Gigabyte HUD program and saw that the load on the GPU is ~ 40%. And progress of WU increased by 0,002% in the second.

I tried the following method. Priority of the process - real time.
Screenshot 1
GPU load was ~ 80%. And progress of WU increased by 0,003%-0.004$ in the second.

Then I returned the priority - below average. GPU load returned to ~ 40%. Progress of WU increased by 0,002% in the second.
Screenshot 2

Then I again tried real time priority.The result was repeated. GPU load ~ 80%.
And progress of WU increased by 0,003%-0.004$ in the second.
Screenshot 3

The screenshots you have to look at Image

---

Another interesting observation. When boinc only uses 2 cores for the calculations on the CPU, and the app 1.19 (atiOpenCL) process is running with realtime priority we have is Screenshot.

1.19 (atiOpenCL) app process uses 25% of the CPU (what is the total load of one core) and GPU load, thus increasing to a maximum of 95%.

My opinion: Too much demand for the CPU, because of what is too low speed of calculation. For comparison, the NVidia 9500 GT sredenee calculation time of one WU 2,5 hours.

Another interesting observation. When boinc only uses 2 cores for the calculations on the CPU, and the app 1.19 (atiOpenCL) process is running with realtime priority we have is Screenshot.

1.19 (atiOpenCL) app process uses 25% of the CPU (what is the total load of one core) and GPU load, thus increasing to a maximum of 95%.

My opinion: Too much demand for the CPU, because of what is too low speed of calculation. For comparison, the NVidia 9500 GT sredenee calculation time of one WU 2,5 hours.

Interesting.

Here is another computor with app. 1.17 http://albert.phys.uwm.edu/result.php?resultid=38871

Her is the same computor with app. 1.19 http://albert.phys.uwm.edu/result.php?resultid=43443

There is great variations in the GPU run time(and the load) between the applications.
And only a smaller variation in CPU run time(and the load)

When I ran my 1.19 i saw a very small usage of my GPU for short periods of time.

As of what I can tell from the people that ran 1.17 they had a higher GPU load then with the new 1.19

I have downgraded to boinc manager 6.12 and I will wait for 6.13 to be the recomended version before running any more tasks.

And I will try the high priority thing you've shoved me when I run the new tasks.

---

Here is another computor with app. 1.17 http://albert.phys.uwm.edu/result.php?resultid=38871

Her is the same computor with app. 1.19 http://albert.phys.uwm.edu/result.php?resultid=43443

If you look in the version 1,17 CPU was used as much as "CPU time 1,382.70". Which led to a greater load of the GPU as there Screenshot and much less runtime WU.(Run time 3,649.92)

If you look in the version 1,19 CPU was used only "CPU time 311.66". Resulting in significant lower loading of the GPU as you say(

When I ran my 1.19 i saw a very small usage of my GPU for short periods of time.

), and much more runtime WU. (Run time 48,329.31)

If I had to pick between 1.17 and 1.19, I would choose 1.17.

If I had to pick between 1.17 and 1.19, I would choose 1.17.[/quote]

Me too.

---

Here is another computor with app. 1.17 http://albert.phys.uwm.edu/result.php?resultid=38871

Her is the same computor with app. 1.19 http://albert.phys.uwm.edu/result.php?resultid=43443

There is great variations in the GPU run time(and the load) between the applications.

Sorry, but that's not caused by the different app versions - I know since this is my machine :-) Have a look again, the latest results of 1.19 are (almost) on par with 1.17 WRT performance. See this one for instance: http://albert.phys.uwm.edu/result.php?resultid=40854

The only difference between those long running task you mentioned is that this time I didn't allow any CPU tasks to run on the same core that the GPU app runs on. This is something we noticed earlier with CUDA apps also. If they don't have enough CPU support, their performance will degrade significantly. Our scheduler specifies that the OpenCL apps need 15-20% of a CPU core and we can just hope that the BOINC client respects that and schedules the host's resources accordingly... We'll keep an eye on this...


Best,
Oliver