Project Update

[tazzduke]

Yeah, I have jumped onboard as well, but am off to a slow start, as I have some DSFL's to clear as well as ALBERT GPU Tasks (Now you can run two tasks at a time without an app_info.xml file, change your account preferences on website) which take two cpu's away from me as well.

Another thing, I have just noticed with BOINC V7.0.25, I have 4 cores being used (I only have a quad), when I started running some POEM and ALBERT tasks on GPU (doing two at a time), my cpu tasks will auto drop back to 2, prior to this release, it was a case of manually setting boinc to only use 2 cores for cpu and allowing 2 free for GPU.  Its just something I noticed. 

I would like a 8 core machine, maybe one day I will get one

Happy crunching and badge hunting
Regards
Tazzduke

[kashi]

Yep, slow start for me too, cache of FAAH won't be completed until tomorrow. I've only finished 3 just to get an idea of how long they take. All 3 took about 4 hours on my Linux VM. None have validated yet, claimed credit is currently about 20% lower than FAAH tasks, I think host scaling is turned off until you have completed 10 tasks.

CFSW tasks use 3-4 times more memory than FAAH ones so I will need to increase the memory allocated to VirtualBox if I want to run all CFSW on Linux. I will give it a go on Windows in a few days and see how it compares.

That's interesting offering the ability to choose 2 concurrent GPU tasks via preferences. I think the difference with 7.0.25 relates to how the maximum ncpus is set. They changed it to prevent idle cores when multi threaded tasks are waiting to run. Then they discovered that change could cause unacceptable slowdown due to CPU overcommitment so they changed it again. I still don't like the 7.0.xx versions at all, not only does the new work fetch cause high priority tasks if one of your projects has a very short deadline but also I experienced a number of "If this happens repeatedly, you may need to reset the project" errors which are a nuisance with CPU applications that have a long checkpoint interval. I thought it was just my computer but I recently saw it mentioned elsewhere.

Some of these new OpenCL GPU apps are very bad with their low GPU load percentage and high CPU usage. POEM requires 3 HT cores to support 4 concurrent tasks and the GPU load percentage is still only 90%. WCG HCC only uses the GPU for a half or less of the task duration and a full CPU core for the rest of the time. Plus it is also CPU bound like POEM, although not nearly as bad but it is still more efficient overall science wise to leave all CPU cores free. Turning off HT and running CPU tasks on 3 physical CPU cores is possibly the best compromise for my setup, but it's pretty crook for badge hunting. Many NVIDIA cards have full or partial busy wait bug so full CPU core is being used for the majority of the task duration. At least the GPU load percentage is better on some of the NVIDIA cards. Busy wait is actually a good idea for WCG, many there are so CPU-centric or new to GPU that they fail to leave a CPU core free to support GPU processing so suffer from CPU overcommitment and the resulting slowdown of both GPU and CPU processing. It doesn't avoid CPU overcommitment slowdown but busy wait ensures the GPU application does not stall due to lack of CPU resources.

[Dataman]

Got a bronze and probably silver tonight. Helps that it is a single quorum project.

[tazzduke]

Glad to hear your input Kashi, and to all others that input through this forum, its how I learn things that happen around in here in BOINC world 

Regards
Tazzduke

[kashi]

Got a bronze and probably silver tonight. Helps that it is a single quorum project.

Congrats with karma on your new badges.

Wish I had known that it's a single quorum project before I downloaded so much work. I've got about 1.5 days worth in my cache and it's all quorum of 2. I've had six validated so far and all but one granted 10-15% less than claimed. Haha, my WCG RAC will drop like a stone with CFSW compared to FAAH.

Strange, when I was doing only a few CFSW tasks at a time the runtime was consistent at about 4 hours but now that I'm doing all CFSW the runtime is varying from 3.1 to 4.1 hours. I've noticed that variability before with other projects such as DNA@Home. Not sure if it's Linux or the virtual machine that causes it. Usually doesn't happen in Windows with no VM, but Windows is often slower.

[Dingo]

Congrats on the new Badges Dataman +2 

[Sean]

With all these new projects is WCG planning to retire any older ones?  ???


+1 Dataman

[tazzduke]

Apologies Dataman, congrats on the new badge and badge colour.  +1 for da karma as well  :penguin:

Regards

[Dataman]

Gold today and doing my "finger & toe" math I will need 8 more days to blue. They are having trouble with stuck wu's on Linux machines. Currently they are limiting Linux to 1 wu/machine.

Just my 2 cents but their GPU project is terribly inefficient. It uses a big chunk of CPU and the GPU starts/stops/starts/stops/ etc. I can't imagine I would crunch it when it goes live.

:aus1: :US

[kashi]

Wondered why I was getting the "No tasks are available" message. Thanks for the heads up.

I'll swap to Windows this afternoon when I finish off the ones in my cache. Linux claimed and granted is going down and down anyway, surely Windows couldn't be any worse. FAAH claims 33 cr/hr and grants 37-45. CFSW is now down to 23.9 cr/hr claimed and 22.8 granted, inconsistent CreditNew strikes again. Haha, I'd just downloaded my first 2 tasks with single replication when the plug was pulled.;D

Running 2 GPU tasks concurrently would possibly fix it up a little, increase GPU utilisation and efficiency plus help to reduce the thermal cycling from the current single task stop/start operation. I made an app_info.xml straight away as soon as the beta was released, it's a corker with lots of <file_name>, <open_name> wrangling. Decided not to test it until the application is released to production though, anonymous platform tasks would probably muck up their testing at this stage. It's got promise if they can do some further development on more recent cards, my HD 5870 takes about 150 seconds to complete a HCC task with HT off which is pretty quick compared to my CPU.

Another karma for your Gold, congrats.

[Mike Mitchell]

Congratulations Dataman!   Plus two. (one now and another in an hour - hope I remember   ).

edit: Yay, I remembered!

[turbo0]

Has there been any word on how GPU cruching will be credited towards badges? Will it just use the same run time system as cpu or be reflective of points?

[danhtruong]

Dataman: Very inefficient GPU processing!

Beta Testing:
WCG Points: 18,086
Results returned: 14
Total Runtime (y:d:h:m:s): 0:001:02:03:28

Repeat: That's 18k WCG points... not BOINC points! So yeh... that's pretty bad.

That's one full day of processing all up on HD 6870's.
If my math is correct, compared to my CPU, that's 4 times faster, but still.... chews up the whole GPU.
Return on GPU power vs points is very low compared to my other Collatz project that I currently run.

[kashi]

OK, for CFSW Windows is faster and gives more consistent runtime than Linux on my box. After losing 3 hours crunching time when my Linux VM ran dry while I was sleeping, I switched to Windows. First 6 tasks on Windows are showing exactly the same progress percentage and should all complete in about 3 hours and 17 minutes.

Dan, I don't understand how your 6870 could take a day to complete 14 GPU tasks. 14 GPU tasks would take about 35 minutes on my 5870, something doesn't add up there. Regardless of that if the credit and (in)efficiency stayed the same as the last batch of betas, WCG GPU would be grudgingly yielding about 10% of what POEM GPU does on my setup.

turbo0, last I saw on the WCG forums before I was attacked and bailed out, the intention is to award badges for GPU work on elapsed time. Currently the beta GPU tasks are being counted as only CPU time, so any GPU time where the CPU is not also contributing is not being counted. A Sapphire badge takes about 3 months on an 8 core computer, but will take 2 years on a GPU if a single concurrent task is run. My GPU would do the the same amount of work that my 8 core computer would process for a Sapphire badge in about 15 days. GPU no  good for badges at WCG, but it avoids upsetting the applecart with CPU only crunchers.

[turbo0]

Thanks for the info guys, got excited when it was announced but might just have to stick to other GPU projects