Nothing wrong, except you were possibly a little impatient!

Thank you Richard, as a matter of fact, if you read my reply to Zalster, I said :

"I don't know...I installed v. 0.44 on top of Boinc, trying to read into the proposed choice that Lunatics gives during the process, but when I started Boinc, the Remaining ( estimated time) were up to 5 times more than when using stock config.
Was I supposed to wait longer and let the system adjust itself ?"

Anyway, I'm willing to give it a try again, and since I use Boinctask, I can report what's happening better...please, stay in the loop !!

Ciao e grazie !
Andrea

Ok Richard, did install Lunatics 0.44.

This was the situation before installing Lunatics :



now, after having installed Lunatics :




As you can see, the estimated Remaining times are preetty scary, but I will let it run to see how the situation is evolving.

Ciao
A.

Andrea, I think you'll find that within a day or 2 it will settle down pretty well, and it might help to actually watch some tasks count down for a while, especially when they are past 60-70% completed, as you watch the countdown timer drop 2-10 seconds every second, and you'll know that it's just in the estimation the program makes while it tried to calibrate itself. Once it's had a chance to 'learn', it will get more accurate, and should be pretty close to it's estimated time when the tasks begin.

---

Andrea, I think you'll find that within a day or 2 it will settle down pretty well, and it might help to actually watch some tasks count down for a while, especially when they are past 60-70% completed, as you watch the countdown timer drop 2-10 seconds every second, and you'll know that it's just in the estimation the program makes while it tried to calibrate itself. Once it's had a chance to 'learn', it will get more accurate, and should be pretty close to it's estimated time when the tasks begin.

Yes, I've noticed the timer's behaviour...I'll let the system adjust and re-check tomorrow.. ;-)
For now, I've noticed that cuda50 re-appeared...I will use BoincTask to compare the result with the the one calculated with opencl_nvidia.
One thing : you said BoincTask can save history in .csv format...how ?

Ciao
A.

It's explained pretty well here. Just click on Extra at the top of BoincTasks, and it brings up the settings menu, where you can configure it.

---

One thing : you said BoincTask can save history in .csv format...how ?

The .cvs files are already created by default, look in:
%APPDATA%\eFMer\BoincTasks\history\

(just Copy/Paste the above line in "Run..." box (Win+R))

---

 


- ALF - "Find out what you don't do well ..... then don't do it!" :)
 

@Al & BilBG,
thank you for the help, I'm already importing it into excel as we speak.
A.

24 hours running lunatics; I know I'm comparing different WUs, but they look of the same family to me ...

HARDWARE:
NVIDIA GPU (device 0) = Gainward Phantom GTX 780ti ( +30 Mhz core OC )
NVIDIA GPU (device 1) = Asus DirectCU II GTX 660ti ( no OC )
NVIDIA GPU (device 2) = Asus DirectCU II GTX 660ti ( no OC )
NVIDIA GPU (device 3) = Asus DirectCU II GTX 660ti ( no OC )
CPU = Intel Core i7-4790K Quad-Core 4 Ghz unlocked ( no OC )


Here's an extract before installing Lunatics 0.44, yesterday morning :





and here is an extract, this morning, with Lunatics running 24h :




One of the first thing I notice is that without Lunatics, the elapsed times were roughly :

11m45s for GTX660ti nVidia Gpus ( opencl_nvidia )
5m50s for GTX780ti nVidia Gpu ( opencl_nvidia )
3h32m for CPU

with Lunatics :

12m4os for GTX660ti nVidia Gpus ( cuda50 )
8m17s for GTX780ti nVidia Gpu ( cuda50 )
1h30m for CPU

As you can see also from the screenshot, it is true that cuda50 app is slightly less performant than opencl_nvidia ( roughly -2.5% for 780ti and -30% for 660ti), but it seems to use a lot less of cpu, so that the cpu crunching get a boost of about +136% ( the WU is crunched in less than a half of the time).

Remember that my app_config.xml is :

<app_config>
<app>
<name>astropulse_v7</name>
<gpu_versions>
<gpu_usage>1</gpu_usage>
<cpu_usage>1</cpu_usage>
</gpu_versions>
</app>
<app>
<name>setiathome_v8</name>
<gpu_versions>
<gpu_usage>1</gpu_usage>
<cpu_usage>1</cpu_usage>
</gpu_versions>
</app>
</app_config>

Also, it seems that the more the GPU is performant, the more it uses the CPU...but this seems obvious.

Now : is getting twice the WU CPU crunched worth keeping than having the GPUs 2.5% and 30% less performant ? ( sorry, I don't know if I expressed myself clearly in english for this... )

A.

As you can see also from the screenshot, it is true that cuda50 app is slightly less performant than opencl_nvidia ( roughly -2.5% for 780ti and -30% for 660ti), but it seems to use a lot less of cpu, so that the cpu crunching get a boost of about +136% ( the WU is crunched in less than a half of the time).

The OpenCL SoG application generally outperforms the CUDA 50 application in their default configuration.
With optimised settings, it significantly outperforms the CUDA50 application, particularly so with the Guppie WUs, however it does require 1 CPU core per GPU WU being crunched to do so.
The increased GPU output more than offsets the loss in CPU crunching output.
Both applications will give more work per hour when running 2-3WUs at a time per GPU than when running just one.


EDIT- the CPU boost is from the use of a more optimised CPU application in the Lunatics installer over the stock one if you select AVX (and your CPU & OS support it).

---

Grant
Darwin NT

The OpenCL SoG application generally outperforms the CUDA 50 application in their default configuration.
With optimised settings, it significantly outperforms the CUDA50 application, particularly so with the Guppie WUs, however it does require 1 CPU core per GPU WU being crunched to do so.
The increased GPU output more than offsets the loss in CPU crunching output.
Both applications will give more work per hour when running 2-3WUs at a time per GPU than when running just one.

EDIT- the CPU boost is from the use of a more optimised CPU application in the Lunatics installer over the stock one if you select AVX (and your CPU & OS support it).

So, what I could do is to change the GPUs behaviour forcing them use OpenCL SoG/sah and let CPU crunching use AVX as in the current environment ?
Is it something to feed to app_info.xml ?

for Lunatics v0.44, the gold standard (highest throughput) seems to be Cuda50 with 2WU/GPU. I got this info from PMs with a few top 1% crunchers who tried SoG when it came out as stock.

Now though, the situation is flipping into SoG's favour as Grant pointed ...but that seems to only apply to top-end/newer nVidia GPUs (I might be wrong).

Before trying Lunatics v0.45 beta3 (and soon beta4), I suggest getting a solid baseline with Cuda50 running 2WU/GPU.

Cheers,
Rob ;-)

---

So, what I could do is to change the GPUs behaviour forcing them use OpenCL SoG/sah and let CPU crunching use AVX as in the current environment ?
Is it something to feed to app_info.xml ?

If running the stock installation, the BOINC manager decides which application to use. Once it decides which is best, it will generally stick with that application- If you run 2, 3 or 4 WUs at a time then the APR (Averge Processing Rate) for that application will drop, and it will start using the other applications again to determine which is better. If you change the number of WUs running at a time for the application you want back to 1, eventually it's APR will be higher than the others & it will be chosen.

Or you can run the Lunatics installer, and you choose the applications that will be used (much easier than manually making your own app_info.xml file).

---

Grant
Darwin NT

@Stubbles69

Didn't know 0.45 was on the way.
Is the upgrade from a previous version smooth, or installing it will "ghost" all the ongoing work ?

I'd like to run 2WU/GPU, did that for a while yesterday, but the amount of system CPU being used was not clear to me.

I am now :

<gpu_usage>1</gpu_usage>
<cpu_usage>1</cpu_usage>

The most logical thing to do for 2 WU should e :

<gpu_usage>0.5</gpu_usage>
<cpu_usage>0.5</cpu_usage>

right ?

@Grant

Yes I thought about using Lunatics, to change app, but on the NV page, the choice is only between various version of cuda ( 23, 32, 42 and 50 ) not between openCL and cuda

Anyway, the tiny amount of CPU used by cuda50, is making me think to remain with that app....

A.

@Grant

Yes I thought about using Lunatics, to change app, but on the NV page, the choice is only between various version of cuda ( 23, 32, 42 and 50 ) not between openCL and cuda

I think I've got it : I just untick any cuda in Lunatics, and then it should use openCL... ;-)

@GazzaVR

So, what I could do is to change the GPUs behaviour forcing them use OpenCL SoG/sah. . .

The most logical thing to do for 2 WU should e :

<gpu_usage>0.5</gpu_usage>
<cpu_usage>0.5</cpu_usage>

right ?

See this explanation.

for Lunatics v0.44, the gold standard (highest throughput) seems to be Cuda50 with 2WU/GPU. I got this info from PMs with a few top 1% crunchers who tried SoG when it came out as stock.
Now though, the situation is flipping into SoG's favour as Grant pointed ...but that seems to only apply to top-end/newer nVidia GPUs (I might be wrong).
Before trying Lunatics v0.45 beta3 (and soon beta4), I suggest getting a solid baseline with Cuda50 running 2WU/GPU.

Before replying to your Qs in reply to quote above, I just realized I made a HUGE mistake; I should have written:

for Lunatics v0.44 and GTX 750 Ti cards, the gold standard...

Sorry for that!

---

@Stubbles69
Didn't know 0.45 was on the way.
Is the upgrade from a previous version smooth, or installing it will "ghost" all the ongoing work ?

That's the beauty of Lunatics v0.45 Beta3: it won't create ghosts when selecting a different GPU app (so you don't have to empty your cache)!
The only thing it doesn't do (from my experience) is:
In Boinc Manager "Tasks" tab, the Application column will still list the previous GPU app name...but if you open up Task Manager, you will see that the GPU is using the selected v0.45 Beta3 GPU app.

---

I think I've got it : I just untick any cuda in Lunatics, and then it should use openCL... ;-)

I started the v0.44 installation to see what the screens looked like (and cancelled the installation before it made any changes to my system).
If you uncheck/untick any NV options, it won't configure any nVidia GPU apps ...whatsoever!
And since the server will dealing with that PC as an "anonymous platform", it won't send the latest SoG stock app.
Therefore, if the above is correct, your nVidia card(s) will be unused by S@H on that PC.

---