Only CUDA new work????


log in

Advanced search

Message boards : Number crunching : Only CUDA new work????

1 · 2 · Next
Author Message
jravin
Send message
Joined: 25 Mar 02
Posts: 930
Credit: 99,174,851
RAC: 84,654
United States
Message 1053934 - Posted: 8 Dec 2010, 15:37:02 UTC

I am getting new work for both my 8-core, 4 GT240 machines now (THANKS to the powers-that-be!) but I am only getting CUDA, even though my machines have asked for both (BOINC 6.10.56). Any idea why this might be? Is the project only generating CUDA WUs for now?
____________

Profile HelliProject donor
Volunteer tester
Avatar
Send message
Joined: 15 Dec 99
Posts: 703
Credit: 87,414,378
RAC: 67,927
Germany
Message 1053935 - Posted: 8 Dec 2010, 15:38:38 UTC - in response to Message 1053934.

Five Minutes ago i got 44 Workunits for my CPU...

Helli
____________

Profile SciManStevProject donor
Volunteer tester
Avatar
Send message
Joined: 20 Jun 99
Posts: 4840
Credit: 81,513,751
RAC: 38,185
United States
Message 1053961 - Posted: 8 Dec 2010, 16:50:37 UTC

Same thing happened to me, but it seems to have settled out now. I even loaded more Einstein last night to keep my chiller from cycling.

Steve
____________
Warning, addicted to SETI crunching!
Crunching as a member of GPU Users Group.
GPUUG Website

Profile BilBg
Volunteer tester
Avatar
Send message
Joined: 27 May 07
Posts: 2679
Credit: 6,053,982
RAC: 3,729
Bulgaria
Message 1053962 - Posted: 8 Dec 2010, 16:51:35 UTC - in response to Message 1053934.
Last modified: 8 Dec 2010, 17:00:51 UTC



I don't think the tasks splitters produce are marked CUDA or CPU -
only after you (or anybody) ask for particular type of tasks
(and successfully get them) they are "marked" accordingly.

(The task (WU) file is the same for CUDA or CPU, only the info about it in client_state.xml is different.
If you look how Reschedulers work to "move" tasks CPU <--> GPU you will see
that only the client_state.xml is changed - not the tasks/WUs files.
)


____________



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

Profile BilBg
Volunteer tester
Avatar
Send message
Joined: 27 May 07
Posts: 2679
Credit: 6,053,982
RAC: 3,729
Bulgaria
Message 1053975 - Posted: 8 Dec 2010, 17:20:02 UTC - in response to Message 1053965.

But if the host asks for both GPU and CPU work (in the same request),
I think it's the scheduler that then marks them as one or the other and decides which work request to fill when they are assigned to a host and sent.


We'll wait for Richard Haselgrove and/or Josef W. Segur to shed some light on how the (server) scheduler decides on that matter ... :)


____________



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

Richard HaselgroveProject donor
Volunteer tester
Send message
Joined: 4 Jul 99
Posts: 8486
Credit: 49,700,956
RAC: 58,081
United Kingdom
Message 1053997 - Posted: 8 Dec 2010, 18:09:12 UTC - in response to Message 1053975.

Sorry, I can only do part of the answer.

Certainly, the WUs themselves are not differentiated: any WU, once split, could be allocated to a CPU, a CUDA card, or any other type of processor for which an application exists. I would imagine that the rule under which a vlar WU is not assigned to a CUDA processor is still in force, but that's an allocation decision, not any fundamental property of the WU itself.

Look at your task list: I'm sure you'll easily be able to find WUs where the two identical replicated tasks (_0 and _1) are allocated, one to CPU and the other to CUDA.

I'll have to leave it to Joe to explain, if he can, why the scheduler faced with a request for both CPU and GPU shows a strong tendency to give up after filling only the GPU part of the request.

Metod, S56RKO
Volunteer tester
Send message
Joined: 27 Sep 02
Posts: 309
Credit: 33,938,411
RAC: 23,760
Slovenia
Message 1053999 - Posted: 8 Dec 2010, 18:18:46 UTC - in response to Message 1053997.

My guess about why scheduler prefers GPU over CPU: GPU will work through task much more efficiently and faster. If scheduler doesn't have enough tasks to fill up both GPU and CPU queues, then it's more than sensible to fill the more effective queue first.

My GPU host can't get enough tasks while some other CPU-only hosts have better luck.
____________
Metod ...

Profile Fred J. Verster
Volunteer tester
Send message
Joined: 21 Apr 04
Posts: 3238
Credit: 31,749,699
RAC: 4,200
Netherlands
Message 1054001 - Posted: 8 Dec 2010, 18:25:10 UTC - in response to Message 1053997.

Bit off topic, haven't seen any Work Units, MB nor AP, lately, but it's still 'early', in Berkeley.
Einstein has given out too much work on 2 QUAD's (Q6600+GTS250&X9650+GTX480) and GPUgrid, that most of their tasks running High Priority, maybe I've to set N.N.T. for Einstein,(looks like 'abuse') :(

But glad new severs are in place and in due time work will come, I hope, atleast.


____________

Josef W. SegurProject donor
Volunteer developer
Volunteer tester
Send message
Joined: 30 Oct 99
Posts: 4241
Credit: 1,046,990
RAC: 302
United States
Message 1054029 - Posted: 8 Dec 2010, 19:06:49 UTC - in response to Message 1053934.

I am getting new work for both my 8-core, 4 GT240 machines now (THANKS to the powers-that-be!) but I am only getting CUDA, even though my machines have asked for both (BOINC 6.10.56). Any idea why this might be? Is the project only generating CUDA WUs for now?

When your host asks for both kinds of work, the Scheduler sends work for the "best" application first and only after that part of the request is fulfilled will it send work for other application[s]. So with a limited amount of tasks available, probably the GPU request is not being fully satisfied and the Scheduler never goes on to send some to the CPU.
Joe

jravin
Send message
Joined: 25 Mar 02
Posts: 930
Credit: 99,174,851
RAC: 84,654
United States
Message 1054037 - Posted: 8 Dec 2010, 19:19:10 UTC

Some further info:

I tried to use the old rescheduler (1.9) to reschedule some GPU -> CPU, but it foumd 0 WUs...is this an artifact of running > 1 Project? It always worked before, but I was only running SETI at the time.

The "new" Rescheduler worked just fine, however.

One question - it doesn't seem to have the option, as the old one did, of rescheduling % of WUs to GPU/CPU...did I miss something?
____________

JohnDKProject donor
Volunteer tester
Avatar
Send message
Joined: 28 May 00
Posts: 842
Credit: 44,050,499
RAC: 73,718
Denmark
Message 1054048 - Posted: 8 Dec 2010, 19:51:04 UTC - in response to Message 1054038.

The old rescheduler does not work with the current WUs and Lunatics apps.

Old rescheduler still works if you edit app_info so it only have the <version_num>608 sections. I'm still using it although there's almost never any vlars send to gpu anymore.

btw be careful editing app_info if you have 609 tasks running, they will disappear from Boinc, finish all gpu work first :)

-BeNt-
Avatar
Send message
Joined: 17 Oct 99
Posts: 1234
Credit: 10,116,112
RAC: 0
United States
Message 1054149 - Posted: 9 Dec 2010, 1:46:18 UTC

At this point in time I'm not going to complain about what work I have, I'm just happy I have work. As I sit now I have a bit over 300 wu's on both my machines. All of it appears to be cuda. Ah well the new quad will have to wait a bit to warm up I guess.
____________
Traveling through space at ~67,000mph!

Proyecto Unity
Avatar
Send message
Joined: 13 Mar 03
Posts: 8
Credit: 4,406,581
RAC: 481
Mexico
Message 1054168 - Posted: 9 Dec 2010, 4:52:43 UTC

I just received an astropulse...now I have cudas , enhanced and astropulse.
____________

dan
Avatar
Send message
Joined: 18 Oct 02
Posts: 392
Credit: 18,182,793
RAC: 12,027
United States
Message 1054177 - Posted: 9 Dec 2010, 5:26:28 UTC

I just got a couple of cpu units marked vlar so the separation of vlars to the cpus is still turned on.
____________

Profile Dirk Broer
Volunteer tester
Avatar
Send message
Joined: 18 Jun 00
Posts: 14
Credit: 3,243,101
RAC: 1,283
British Virgin Islands
Message 1054230 - Posted: 9 Dec 2010, 12:57:19 UTC

Strange, since I let my GPU's also work at PrimeGrid I do not get *any* CUDA work?
____________

1 · 2 · Next

Message boards : Number crunching : Only CUDA new work????

Copyright © 2014 University of California