Posts by Mutiny32*
log in
1) Message boards : Number crunching : CUDA 2.04 Application Hints, Tips, Performance, and Issues (Message 9691)
Posted 2681 days ago by Profile Mutiny32*
Fermi should run about the same speed though.
2) Message boards : Number crunching : CUDA 2.04 Application Hints, Tips, Performance, and Issues (Message 9682)
Posted 2681 days ago by Profile Mutiny32*
Does it work with 64 bit windows ?


Yes. Regardless of the operating system and even though the drivers are specific to 32 or 64-bit operating systems, the GPUs are still only 32-bit so the 32-bit app works really well on them. There is no benefit to having a 64-bit GPU app since it will run at exactly the same speed as the 32-bit app because the GPU really is 32-bit.

Actually Fermi can do FP64 operations as fast as FP32 operations, as it is designed to do FP64 natively. As for compiling, I'm not 100% on this, but I believe you can only compile 32-bit CUDA apps on Windows, but on Linux, you can compile 32 or 64-bit CUDA apps.
3) Message boards : Number crunching : CUDA 2.04 Application Hints, Tips, Performance, and Issues (Message 9681)
Posted 2681 days ago by Profile Mutiny32*
I find that
<type>CUDA</type>
<count>0.333333</count>


In the app_info.xml is the most efficient setting in the 2.05 app_info.xml for Windows on my GTX460 under L13 I8 S1 as well as the lowest CPU time used, but it has an unfortunate side-effect of the app not fetching more WUs when it has gone through all of them. I finally figured out that the CUDA device count has to be set at 1 or else it simply won't fetch any more WUs. This means I set my work fetch to around 7 days extra so I don't have to go back and change it to 1, restart BOINC, let it download the WUs, exit BOINC again, change it back to 0.333333, and finally start BOINC back up.

I wish there were some sort of workaround.

I also have my FLOPS set to 7.5797e11 as measured under CUDA-Z, it shows with Windows running including Aero, 757.97 GFLOPS is the average for single-precision floating-point operations. The official specs are 907 GFLOPS for the card, but with Windows/Aero/whatever else running, I'm using a realistic number here. Am I right in changing that number to be accurate?
4) Message boards : Number crunching : Linux optimized app nice level doesn't play nice. (Message 9531)
Posted 2691 days ago by Profile Mutiny32*
Can I at least get an acknowledgement? the guys at AQUA were surprised to find out that Collatz ran at a nice of 10.

What's the point in having it at this nice rating? It doesn't make the app useful in Linux because you have to change the nice level manually. Is there some sort of CUDA tweaking guideline that says to do this?

I'd imagine the same is going to happen for the new OSX version as well, considering it is 'nix based.


Collatz default is to run at nice 0, so i don't know how you managed to get it running at 10.

It's most likely caused by your setup. Try using a proper linux derivate like Fedora or OpenSuSE. (unless you're into beeing a guinea pig for Ubuntus pre alpha software junk).



I run Mint x64. Putting it at a nice rating like that doesn't play nice with other apps. By the way, Ubuntu is now based off of Debian testing, so you can throw your argument about pre-alpha out the door.

But hey, whatever, I'll just halt my WU times that run around 700 seconds until I can figure out how to make it act like it should.
5) Message boards : Number crunching : Linux optimized app nice level doesn't play nice. (Message 9529)
Posted 2691 days ago by Profile Mutiny32*
Can I at least get an acknowledgement? the guys at AQUA were surprised to find out that Collatz ran at a nice of 10.

What's the point in having it at this nice rating? It doesn't make the app useful in Linux because you have to change the nice level manually. Is there some sort of CUDA tweaking guideline that says to do this?

I'd imagine the same is going to happen for the new OSX version as well, considering it is 'nix based.
6) Message boards : Number crunching : Linux optimized app nice level doesn't play nice. (Message 9498)
Posted 2693 days ago by Profile Mutiny32*
Ever since I installed the Linux 2.04 Collatz x86-64 CUDA 3.1 optimized app, I've been having a troubling issue with other project CPU WUs; specifically those from projects that take a very long time to complete like AQUA and ClimatePrediction, causing a TON of wasted CPU time to because they kept starting WUs over. I spent quite a bit of time troubleshooting the issue and finally deducted a correlation with the collatz GPU app and the other apps running at the same time.

Then I noticed it. Collatz runs at a nice level of 10. This contrasts with EVERY app I run having a nice level of 19. The same goes for the Windows CUDA 3.1 collatz app, but with less problems due to the priorities not being so different; nontheless the collatz app still being selfish and using a higher priority than all other apps.

I understand that this is an optimized app in both cases, but I don't see ANY reason for the default nice level to be changed to 10 from 19. This is wasting a TON of my CPU time on WUs that are earning me no credit because they keep getting killed by the OS prioritizing collatz. I would advise whoever codes and compiles these apps to stop screwing with nice levels or at least give us an option to make it play nice in app_info.xml.

BTW, I noticed that the Linux cmdline options of 13, 8, and 0 to bring the same system to a virtual halt versus the same switches on W7 x64 that hardly dents UI responsiveness. Is that an X.org thing?
7) Message boards : Number crunching : CUDA 3.1 application 2.05 faster (Message 9386)
Posted 2699 days ago by Profile Mutiny32*
You have to use newer drivers in order to use the new CUDa app. go to laptopvideo2go.com and grab the latest drivers like 259.32 and install them for your platform.
8) Message boards : Number crunching : CUDA 3.1 application 2.05 faster (Message 9383)
Posted 2699 days ago by Profile Mutiny32*
On an 8600GT, it is not friendly toward GUI resonsiveness. I'm playing with the L number and the avg_ncpus/max_ncpus to see if I can't get usability back on Win7 x64.
9) Message boards : News : CUDA 3.1 Application Available for Windows (Message 9382)
Posted 2699 days ago by Profile Mutiny32*
Err, gotta bit of a question. Two actually. Is the executable 64-bit? It appears to be 32 by the name of it. Second, you have the following in the app_info.xml


<cmdline>L13 I8 S0</cmdline>
<avg_ncpus>0.013</avg_ncpus>
<max_ncpus>1</max_ncpus>


That cmdline L13 is NOT GUI friendly in terms of responsiveness or even usability.

The avg_ncpus and max_ncpus different from that of the 3.1 CUDA app for my linux machine, which has 0.03 for both values. It appears to be different than the cuda23 windows app as well, because it always used a max of 0.03 CPU. This in its stock form shows up as 0.01 CPU usage. Did you mean for it to be 0.13 perhaps? Or did you find something else out that caused this change?




Main page · Your account · Message boards


Copyright © 2018 Jon Sonntag; All rights reserved.