Text only | Text with Images

BOINC@AUSTRALIA FORUM

Retired BOINC Projects => DNETC@HOME => Topic started by: Dingo on April 03, 2010, 01:47:45 PM

Title: New Project -
Post by: Dingo on April 03, 2010, 01:47:45 PM
I mistakenly deleted the post instead of moving the posts.  Me BAD.   :cry2: Lucky I had a screen open with the posts in it so I could copy and paist.

Sorry for the inconvienience but carry on crunching.


kashi
Hero Member
*****

Aussie Karma: +157/-0
Title: Re: New Project -
Post by: Dingo on April 03, 2010, 01:51:05 PM
It looks like they also have the CUDA app running as well.  There still seems to be a problem  (http://dnetc.net/forum_thread.php?id=13)if you are running an older card with less than 384 MB ?  :shock

I started a couple with mt GTX 275 and they are crunching, but not have completed as yet.
Title: Re: New Project -
Post by: kashi on April 03, 2010, 03:13:14 PM
Looks like it works, but in a strange way. After finishing one packet, the CUDA application assesses that the remaining 11 packets will take too long to complete and so terminates at that stage. It validates OK and grants the same proportional amount of credit, that is one twelfth of the credit that is granted for the 12 packets completed on an ATI card.

This is what I was referring to previously when I said that I wasn't sure the CUDA application could handle the new longer tasks. After looking at the stderr out section of task details I realised why CUDA tasks are granting less credit. I assume this is by design and is a way of getting around the problem of excessively long running CUDA tasks with no progress bar.

I don't know how the credit rate compares to other CUDA projects, perhaps not too favourably seeing how long NVIDIA cards take to process packets and tasks on this project.
Title: Re: New Project -
Post by: Dingo on April 03, 2010, 06:03:51 PM
My GTX 275 is doing Wu in about 443 seconds and getting 120.80 credit which equates to almost 982 credits per hour.  It uses more CPU power than most of the GPU projects at 0.40 % CPU where the others like MilkyWay and Collatz only use 0.05 or 0.04 % CPU.

On the same machine Collatz gets about 1762 credits per hour and MilkyWay 1321 credits per hour.   So you get less credit than the other GPU projects and you also loose 40% of one CPU that cannot crunch CPU projects. 

I am still leaving it running with with the other GPU projects for a while while it is new.
Title: Re: New Project -
Post by: kashi on April 03, 2010, 07:33:13 PM
Yep I thought NVIDIA times looked a bit slow.

I agree the CUDA application is using too much CPU time.

Although it is close in your GTX 275 results, the CPU percentage shown in BOINC Manager for GPU tasks doesn't indicate how much CPU time is actually used. The value is used to allocate cores to support GPU crunching. This project automatically uses multiple GPUs concurrently, so 2 GPUs still only use 0.4 CPU core working together on a single task. However running MilkyWay for example if you had 2 GPUs processing a task each and you set the ncpus value to 0.4 cores then together they would use 0.8 of a core so only 3 CPU cores on a quad would be available to process other projects.

That's the original theory of it anyway, although sometimes it doesn't always work that way with some current applications such as AQUA. I can now process AQUA tasks on all 8 cores while doing MilkyWay even though I have a full CPU core allocated to support GPU crunching. If I do other projects like QuantumFIRE only 7 cores of it will run while crunching MilkyWay.

From the look of your results the CUDA application on your GTX 275 uses about 186-189 seconds of CPU time to complete one packet of a standard length 12 packet task.

Contrast this with your ATI card which is also shown by BOINC as using 0.4 of a CPU core but processes all 12 packets and only uses just over a second of CPU time.
Title: Re: New Project -
Post by: BF on April 03, 2010, 09:08:54 PM
Ok, have crunched this project for 12 hours and picked up 100,000 credits already.. for my cards, that's a tad less than MilkyWay and somewhat more than Collatz when calculating cr/hr..

Will stick with it for a few days to see how project stability compares with MilkyWay..

Great to finally have a third option for ATI cards!  :dance:
Title: Re: New Project -
Post by: LawryB on April 04, 2010, 08:15:35 AM

My 5870 on 800/900Mhz fan 60 temp 72 activity 98. 

Now seems to have settled into doing WUs as follows

dnetc_gpu_38xxx_x in 200 secs  (these the old WUs ?????)

dnetc_gpu_11xxxx_x in 480-495 secs.

Went to make a coffee and it racked up 30K (OK slight exageration).  Only had one computation error.

My pair of 4770s on my I7 with 16GB RAM went ape.  A single WU pinched both GPUs.   Little or no activity showed up on the ATI console and WUs were taking over 1.5 hours when they did complete.

I'll swap them for a couple of 5770s later this day (when  >:D allows me to stop working in the yard) and let you know.

I do hope the Milky Way hackers don't get active on this one.
Title: Re: New Project -
Post by: kashi on April 04, 2010, 09:57:03 AM
As you said the shorter WUs are probably old ones, I think they are reissues (-1 -2 etc. at the end of the name).

Yes a few people on the DNETC forums have been having trouble with multiple GPUs although it works well for others. The troubles include very poor performance or the tasks freezing for hours without progressing. The performance issues not the freezing may be the model of cards used, perhaps the more recent ones with a larger amount of memory can handle running one task on 2 cards much better.

If you look at your 3 results with dual cards the only result that has stderr out details shows that the application stopped and started a few times. Your 5870 gobbles up WUs in one sitting however (no restarts).

There is a L'AF member who posted on the forum recently that he had managed to get dual cards to run one task each on his 4850s and this cured his problems. Couldn't understand exactly how he did it. I think he means you start BOINC Manager with one card "plugged" (connected to multiple input monitor or use dummy plug) and then download DNETC tasks and they will be marked to run on a single GPU after you plug the second card again. Don't know what would happen if you downloaded more tasks though. Wouldn't want to have to stop and unplug and replug your 2nd card every 40 tasks.  :(

BTW this project is Windows only, I noticed a member had attached a computer running Darwin.
Title: Re: New Project -
Post by: LawryB on April 04, 2010, 11:49:45 AM
Thanks kashi.  Makes a bit more sense now.  Wonder what my 5970 will do with these?  When I get it that is.
Title: Re: New Project -
Post by: kashi on April 04, 2010, 02:07:33 PM
Ohh a 5970, I've had my eye on those for a while now. The Amaze branded ones were down to from $800 until the Fermi reviews came out, then they popped back up to from $900. I'm not sure one would fit in my case, haven't measured. I was more concerned with noise, I would like to get a reference one and whack an Accelero XTREME cooler on it but they won't be available until May or so, that's if any stock reaches Australia.

DNETC will process one task on both cores of a 5970 same as it does for 2 separate GPUs in a box.

In other news, the team has just cracked 1 million in this project and is currently 12th in the team rankings. It may not last but it's not a bad start.  :dance:
Title: Re: New Project -
Post by: kashi on April 04, 2010, 09:14:55 PM
OK, I did a Project Overview, could an administrator please sticky/pin the post and lock that thread please. Thanks.
Title: Re: New Project -
Post by: Dingo on April 04, 2010, 09:42:21 PM
Done,  Thanks great job  :congrats
Title: Re: New Project -
Post by: Zed on April 05, 2010, 12:12:51 AM
Hmmmmm

I have 2 Gtx cards running in 1 box and thought i would give this a go. it seems this project wont use the 2nd card and it wont run along side collatz.

Tried to run it on my ati card and all it did was error. not gonna bother with this project
Title: Re: New Project -
Post by: LoneWolf_53 on April 05, 2010, 01:59:46 AM
I've put one card on it.

Too bad they don't have a Linux provision.
Title: Re: New Project -
Post by: dyeman on April 05, 2010, 01:22:26 PM
I'm running it together with Collatz on a 4830 (Win XP) and it seems to be OK.  Will drain the Collatz and give it a bit of a run.
Title: Re: New Project -
Post by: kashi on April 06, 2010, 08:47:26 AM
Longer wus for Radeon 5k series
Server now send longer wus for Radeons 5xxx (24 packets x 32 units) 5 Apr 2010 22:09:30 UTC

Good timing there seeing as the Radeon 58xx and 5970 are currently producing results that are not accurate enough on MilkyWay. Wooshka, did a new 24 packet one. The standard long tasks for Radeon 5xxx series grant 2,982.25 credits each.

Edit: Now increased further to 48 packets x 16 units
Title: Re: New Project -
Post by: Furlozza on April 06, 2010, 02:00:58 PM
The ones I am doing grant 4800+ crs.... on my 5870 for just 24-25 mins work.... albeit at slightly higher temp and fan speed than MW or Collatz. Is actually using the GPU 97% of the time, according to ATI's control panel.
Title: Re: New Project -
Post by: kashi on April 06, 2010, 03:17:52 PM
Yes the ones paying above 4,000 are 48 packet WUs developed for ATI 5xxx models. They released the 24 packet ATI 5k ones first which was the first longer one I completed this morning as mentioned in my previous post, then 20 minutes later decided to double them again to 48 packets. There were still 24 packet WUs on the server though so it took a while to issue those and for me to work through my cache until I reached the 48 packet ones. I 'm doing my third one of these monsters now. The intention is for all the dnetc_GPU_ATI_5k_xxxx series WUs to be 48 packet size from now on.

Yes this is the hottest running of the ATI GPU projects for me too. I hope you have your GPU memory speed reduced for this project.
Title: Re: New Project -
Post by: Furlozza on April 06, 2010, 07:05:11 PM
For the moment, it is running on factory speeds ...... mainly because have yet to figure out how to change it LMAO
Title: Re: New Project -
Post by: kashi on April 06, 2010, 09:40:52 PM
Ah, lucky you've got the Vapor-X version then, if you tried DNETC on a reference cooler 5870 with default memory speed you would probably have one very hot card.
I use AMD GPU Clock Tool for all adjustments to overcome the overclocking limit in Catalyst Control Centre but perhaps just enabling Overdrive in CCC and moving the memory slider to the left would do it for you.

No pressure, you are of course free to do whatever works for you, it was meant only as a helpful suggestion not a command.  ;D
Even if you are comfortable with the temperature and  noise level, you can save a bit of power by reducing the memory speed. Lower power bill = more beer or travel!

I noticed that some of the new ATI 5K WUs are now completing all 48 packets and sometimes even 49 packets. The credit for some of these is nearly 6,000 each.
Title: Re: New Project -
Post by: kashi on April 16, 2010, 04:18:00 AM
News

New ATI app version

New app should finally fix the problem with 'infinite workunits'. Wrapper check wu every 20 minutes if detect 'infinite wu', It'll abort this task. Also fixed problem with "Your NVIDIA GPU has insufficient memory (need 384MB)", now every card with 128 MB RAM and more will work.

EDIT: App 1.15 have bug on some types of card. You can abort all 1.15 wus and download 1.16. 15 Apr 2010 7:44:21 UTC
Text only | Text with Images