ABP1 CUDA applications

Michael Goetz
Joined: 11 Feb 05
Posts: 15
Credit: 170,028
RAC: 0

RE: Hi all! Which apps

Message 95647 in response to message 95646

Quote:
Hi all!
Which apps should I uncheck in "Run only the selected applications" section of preferences edition page to receive only cuda-related tasks, and stop to receive pure-CPU WUs?

The applications you choose aren't important; just leave them all checked.

At the top of that section, there's two boxes "Use CPU" and "Use NVIDIA GPU". Check the NVIDIA GPU box and uncheck the CPU box. As noted on the form, you'll need BOINC client 6.10+ for that to work.

That being said... I'd strongly recommend just the opposite -- turn off the GPU and only accept CPU work from Einstein, at least for now. At least if there's any other GPU project you want to do work on. At this time, the Einstein CUDA application hardly uses the GPU at all -- it *barely* runs any faster than the pure CPU app (about 33%), but it uses up both a full CPU core and the GPU. By turning on GPU tasks you're preventing your computer from doing other meaningful work on the GPU, not speeding up Einstein a whole lot, and still using a full CPU core to do the Einstein crunching.

Want to find one of the largest known primes? Try PrimeGrid. Or help cure disease at WCG.

Gundolf Jahn
Gundolf Jahn
Joined: 1 Mar 05
Posts: 1,079
Credit: 341,280
RAC: 0

RE: Hi all! Which apps

Message 95648 in response to message 95646

Quote:

Hi all!

Which apps should I uncheck in "Run only the selected applications" section of preferences edition page to receive only cuda-related tasks, and stop to receive pure-CPU WUs?


You can't. You'd have to uncheck the GW tasks, and that's not possible.

Quote:
btw, somewhy I can't uncheck Hierarchical S5 all-sky GW search #5 and #6....


That's a feature, not a bug ;-)

Gruß,
Gundolf
[edit]This time, I was too slow (again :-)[/edit]

Computer sind nicht alles im Leben. (Kleiner Scherz)

k6xt
k6xt
Joined: 11 Sep 05
Posts: 7
Credit: 22,445,759
RAC: 365

RE: RE: Hi

Message 95649 in response to message 95647

Quote:
Quote:
Hi all!
[Snip]
That being said... I'd strongly recommend just the opposite -- turn off the GPU and only accept CPU work from Einstein, at least for now. At least if there's any other GPU project you want to do work on. At this time, the Einstein CUDA application hardly uses the GPU at all -- it *barely* runs any faster than the pure CPU app (about 33%), but it uses up both a full CPU core and the GPU. By turning on GPU tasks you're preventing your computer from doing other meaningful work on the GPU, not speeding up Einstein a whole lot, and still using a full CPU core to do the Einstein crunching.

Some of these comments are really, and unnecessarily, snippy (not including the above). Just turn off GPU at Einstein. After reading this thread I turned off GPU. Immediately the GPU switched back to SETI as I expected. I also had to reset the Einstein project because all I had were CUDA WU.

Mike Hewson
Mike Hewson
Moderator
Joined: 1 Dec 05
Posts: 6,020
Credit: 102,258,199
RAC: 147,875

Just in case of

Just in case of misconception. E@H is about GW detection. The ABP is an optional sideline if you're keen. That's why there isn't a GW opt out. This is the policy of the project ( no doubt because GW detection is the project ). And no, that's not likely to be re-defined. :-)

Cheers, Mike.

I have made this letter longer than usual because I lack the time to make it shorter. Blaise Pascal

cristipurdel
cristipurdel
Joined: 19 Jul 07
Posts: 23
Credit: 10,900,412
RAC: 25

Any news on ABP2 or maybe an

Any news on ABP2 or maybe an ATI application?

DanNeely
DanNeely
Joined: 4 Sep 05
Posts: 1,273
Credit: 1,279,932,010
RAC: 1,070,816

RE: Any news on ABP2 or

Message 95652 in response to message 95651

Quote:
Any news on ABP2 or maybe an ATI application?

An improved ABP app that does most of the calculations in single precision and which will allow offloading most of the work to the GPU is in work; ETA when it's done.

In the past ATI's declined to provide any assistance in developing a GPU app; so an ATI app isn't likely anytime soon.

rroonnaalldd
rroonnaalldd
Joined: 12 Dec 05
Posts: 116
Credit: 537,221
RAC: 0

RE: In the past ATI's

Message 95653 in response to message 95652

Quote:
In the past ATI's declined to provide any assistance in developing a GPU app; so an ATI app isn't likely anytime soon.

Maybe the user "gipsel" from milkyway can help in developing a GPU app.

hotze33
hotze33
Joined: 10 Nov 04
Posts: 100
Credit: 293,019,576
RAC: 32

I recently switch from the

I recently switch from the beta app to this new official one. I noticed some discontinuity between the actual used processing time and the time shown in the CPU Time column for a wu.
In the beta app it was roughly 4 hours real processing time and time reported
(149011603 ) and every wu took the same amount of time.
Now with the new app the real time is up to 18k s and reported times are much lower more like 13k s.
Also the variance is much more pronounced. I have short ABP units which only take 10.5k s and claim less credit (70) but get also 250cr like the ones which take 14k s and claim 90 cr. The beta app usually claimed 90 cr and took also 14k s.
Is there a new type of wu out?

Gary Roberts
Gary Roberts
Moderator
Joined: 9 Feb 05
Posts: 4,794
Credit: 27,229,535,087
RAC: 35,094,555

RE: I recently switch from

Message 95655 in response to message 95654

Quote:
I recently switch from the beta app to this new official one. I noticed some discontinuity between the actual used processing time and the time shown in the CPU Time column for a wu.


I don't have an NVIDIA GPU so some of the following comments are guesses about what might be happening.

With modern BOINCs, I think that the value reported in the CPU time column of BOINC manager is wall clock time rather than actual CPU time. If a task was waiting for the GPU, the wall clock time would keep ticking but the CPU time wouldn't.

In your tasks list, all the tasks showing as completed on Dec 6 and Dec 7 are ABP1 tasks and all seem to show use of the GPU - I didn't check every single one. As you have a quad core - 4 simultaneous tasks - and 1 GPU, I presume there may be quite a few times where some tasks are waiting for another to release the GPU. There didn't seem to be any GW tasks crunched during those 2 days.

Quote:
In the beta app it was roughly 4 hours real processing time and time reported
(149011603 ) and every wu took the same amount of time.


Since the beta test used AP, presumably your app_info.xml specified both the GW app and the APB1 app. Perhaps you had plenty of GW tasks running and perhaps the APB1 tasks weren't competing for the GPU. I didn't really spend time analysing the older parts of your task list but I did notice a lot more GW tasks there.

Quote:
Now with the new app the real time is up to 18k s and reported times are much lower more like 13k s.


I think you can discount the 18K values as being wall clock time. The real variation seems to be largely in the 12-14K range. My own impressions are that even on CPU only crunching, there is some variability that seems to come in batches from time to time. It's only an impression as I haven't looked in any detail.

Quote:
Also the variance is much more pronounced. I have short ABP units which only take 10.5k s and claim less credit (70) but get also 250cr like the ones which take 14k s and claim 90 cr. The beta app usually claimed 90 cr and took also 14k s.


I don't know why you got a value as low as 10.5K. Perhaps there is more variation in what the GPU use can save and perhaps that was a task with little or no GPU contention. I'm only guessing.

The variation in credit claim is quite normal since the claim by the client is based on benchmarks and CPU time. The variation in true CPU time will give a variation in claim. The credits are granted server side and completely ignore the claim.

Quote:
Is there a new type of wu out?


Not that I'm aware of.

Cheers,
Gary.

Bikeman (Heinz-Bernd Eggenstein)
Bikeman (Heinz-...
Moderator
Joined: 28 Aug 06
Posts: 3,515
Credit: 426,683,895
RAC: 191,845

RE: I don't know why you

Message 95656 in response to message 95655

Quote:

I don't know why you got a value as low as 10.5K. Perhaps there is more variation in what the GPU use can save and perhaps that was a task with little or no GPU contention. I'm only guessing.

Indeed, this is curious. Not all WUs are alike, there is some data dependency in the WUs' runtime, but this is the most noticeable I've ever seen. I suggest to wait if it validates and then see if the wingman also spent a less-than-average runtime on it.

CU
Bikeman

Comment viewing options

Select your preferred way to display the comments and click "Save settings" to activate your changes.