Advanced search

Message boards : Number crunching : Granted credits

Author Message
marsinph
Send message
Joined: 11 Feb 18
Posts: 17
Credit: 34,951,075
RAC: 877,701
Level
Val
Scientific publications
wat
Message 50361 - Posted: 1 Sep 2018 | 17:09:01 UTC

Hello,
Start situation : same host (46603). Same day of WU,always long run WUV9.22 cuda 80. I repeat again, on the same host.
One WU run 50,742 sec for credit of 181,050 (ratio : 3.56 cr/second)
A other run 32,598 sec credit : 110,400 (ratio : 3.38 cr / sec)
OK so fa it is more or less the same

A other run 24,129 sec credit : 63,750. Ratio : 2.64cr/sec

I repeat, I ompare three WU on the same host.
Between hot it can change. Here not it is the same host.



Can someone explain me the ratio difference ???
All my host are visible, als WU ???

____________

mmonnin
Send message
Joined: 2 Jul 16
Posts: 255
Credit: 647,700,389
RAC: 65
Level
Lys
Scientific publications
wat
Message 50362 - Posted: 1 Sep 2018 | 18:13:41 UTC

They are different apps. The short ones probably utilizes the GPU less efficiently.

Richard Haselgrove
Send message
Joined: 11 Jul 09
Posts: 902
Credit: 2,140,463,345
RAC: 515,471
Level
Phe
Scientific publications
watwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwat
Message 50363 - Posted: 1 Sep 2018 | 18:14:06 UTC - in response to Message 50361.

Both BOINC and GPUGrid broke the equivalence between credits and processing power (GFlops) several years ago. Credit is now just a dimensionless number, enabling comparison between hosts and users within each separate project.

Here at GPUGrid, we went through a phase where there was a fairly high probability that tasks would fail mid-run. And tasks which fail get no credit at all.

The longer a task runs before failure, the greater the loss of potential credit. And the longer a task is planned to run, the higher the average loss per failed task. For that reason, the administrators here decided to grant a higher rate of credit per hour for longer running tasks, to provide some compensation for the observed failures.

Your figures illustrate perfectly - higher rate for longer tasks - that this policy still applies and is working as designed.

Richard Haselgrove
Send message
Joined: 11 Jul 09
Posts: 902
Credit: 2,140,463,345
RAC: 515,471
Level
Phe
Scientific publications
watwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwat
Message 50364 - Posted: 1 Sep 2018 | 18:14:54 UTC - in response to Message 50362.

They are different apps. The short ones probably utilizes the GPU less efficiently.

Same apps, different tasks.

marsinph
Send message
Joined: 11 Feb 18
Posts: 17
Credit: 34,951,075
RAC: 877,701
Level
Val
Scientific publications
wat
Message 50369 - Posted: 2 Sep 2018 | 7:25:11 UTC

Hello,
Once again I repeat : I compare differents tasks from the same apps (long run...) on the same host (46603) !!!
I look the credits on my stat page here ongpugrid.

I NOT compare host/apps/...
Only same apps on same host
Waiting explanation, best regards

tullio
Send message
Joined: 8 May 18
Posts: 157
Credit: 37,709,395
RAC: 12,901
Level
Val
Scientific publications
wat
Message 50370 - Posted: 2 Sep 2018 | 7:51:53 UTC - in response to Message 50369.

I am running 5 BOINC projects. Each of them has a different method for assigning credits. My only opinion is that the more they are the less they are worth, exactly like money.
Tullio

Richard Haselgrove
Send message
Joined: 11 Jul 09
Posts: 902
Credit: 2,140,463,345
RAC: 515,471
Level
Phe
Scientific publications
watwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwat
Message 50371 - Posted: 2 Sep 2018 | 8:57:00 UTC - in response to Message 50369.

Hello,
Once again I repeat : I compare differents tasks from the same apps (long run...) on the same host (46603) !!!
I look the credits on my stat page here ongpugrid.

I NOT compare host/apps/...
Only same apps on same host
Waiting explanation, best regards

I repeat my answer. It is a deliberate policy applying to this project only.

marsinph
Send message
Joined: 11 Feb 18
Posts: 17
Credit: 34,951,075
RAC: 877,701
Level
Val
Scientific publications
wat
Message 50372 - Posted: 2 Sep 2018 | 13:31:04 UTC - in response to Message 50370.

I am running 5 BOINC projects. Each of them has a different method for assigning credits. My only opinion is that the more they are the less they are worth, exactly like money.
Tullio



Hello Tulio. Please read all !!!
For the third time, I repeat : same project, same app, same host, only the WU is different and coming from the same batch file !!!
Both credit calulated by GPUGrid

Richard Haselgrove
Send message
Joined: 11 Jul 09
Posts: 902
Credit: 2,140,463,345
RAC: 515,471
Level
Phe
Scientific publications
watwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwat
Message 50373 - Posted: 2 Sep 2018 | 13:47:04 UTC - in response to Message 50372.

Both credit calculated by GPUGrid

Yes - but calculated by GPUGrid staff, not by the BOINC software.

Look at the three tasks you started this thread with.

ADRIA_FOLDADAH2: Credit 181,050.00
PABLO_2IDP: Credit 110,400.00
ADRIA_FOLDUCB: Credit 63,750.00

Each task type gets a fixed credit, but different task types get different credits - as I think you've said already. The staff decide that.

marsinph
Send message
Joined: 11 Feb 18
Posts: 17
Credit: 34,951,075
RAC: 877,701
Level
Val
Scientific publications
wat
Message 50374 - Posted: 2 Sep 2018 | 13:49:12 UTC - in response to Message 50371.

Hello,
Once again I repeat : I compare differents tasks from the same apps (long run...) on the same host (46603) !!!
I look the credits on my stat page here ongpugrid.

I NOT compare host/apps/...
Only same apps on same host
Waiting explanation, best regards

I repeat my answer. It is a deliberate policy applying to this project only.



@Richard.
In wich language I need to write ?
I repeat : same project (GPU), same host (46603), same Apps (long time...),
I know the credit change between projects.
I ompare what is possible.
To compare :
Take a car. You need to drive 100km.
The first time you drive at 100km/h, so one hour driving needed. Let us say 10 liters gasoline required.
The second time, you again, in the same car, also 100km on the same road. Sometime you drve faster (more gasoline needed), sometime slower (less gasoline needed). Nomally about the same as first race (let us says about 5% differences.
Then a third race, perfect the same conditions as the first. But you need 12 liters gasoline : 20% more


So why there is so big difference with the third race, same road, car, driver, weather,......? All is perfect the same !!!



PappaLitto
Send message
Joined: 21 Mar 16
Posts: 502
Credit: 4,274,088,001
RAC: 1,808,270
Level
Arg
Scientific publications
watwatwat
Message 50375 - Posted: 2 Sep 2018 | 13:57:04 UTC

@marsinph

I think I may have the final answer to this. I am honestly not sure how the GPU Work Unit credits are calculated, either by the staff or automatically based on total FLOPS. If it is the latter I think you will notice a lack of GPU usage on certain WUs. This directly affects the total WU time and therefore credit/second. Because the app sends all the data over the PCIe bus to the CPU for double precision compute, if you single threaded CPU speed is lacking or you have a less than optimal PCIe configuration, this can lead to low GPU usage.

If the WU is a very large number of molecules, very little CPU usage is required as compared to GPU usage and most likely your credit/second will be higher.

Richard Haselgrove
Send message
Joined: 11 Jul 09
Posts: 902
Credit: 2,140,463,345
RAC: 515,471
Level
Phe
Scientific publications
watwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwat
Message 50376 - Posted: 2 Sep 2018 | 14:42:19 UTC - in response to Message 50375.

I am honestly not sure how the GPU Work Unit credits are calculated, either by the staff or automatically based on total FLOPS.

At this project - unusually - the credits are calculated by the staff.

Look at your total list of GPU tasks. The credit for every single one of them (he sticks his neck out, risking egg on face!) is an exact multiple of 50.00. The chances of that happening by floating point arithmetic, whether you're counting flops or microseconds, are minuscule.

tullio
Send message
Joined: 8 May 18
Posts: 157
Credit: 37,709,395
RAC: 12,901
Level
Val
Scientific publications
wat
Message 50377 - Posted: 2 Sep 2018 | 14:46:07 UTC - in response to Message 50375.
Last modified: 2 Sep 2018 | 14:47:39 UTC

I don't know how many molecules are in the GPU tasks but Toni wrote that the CPU tasks treat only one molecule. I have tried to understand in the "output" directory of stderr.txt what the program is doing but it is beyond my knowledge of computer chemistry. I only remember that SCF stands for Self Consistent Field and DFT for Density Functional Theory.
Tullio
____________

Profile Retvari Zoltan
Avatar
Send message
Joined: 20 Jan 09
Posts: 2031
Credit: 14,709,489,669
RAC: 1,510,420
Level
Trp
Scientific publications
watwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwat
Message 50378 - Posted: 2 Sep 2018 | 19:30:19 UTC - in response to Message 50374.
Last modified: 2 Sep 2018 | 19:42:18 UTC

Here at GPUGrid the same app does very different simulations between different workunit batches (they are named differently).
The granted credits are set manually by the GPUGrid staff for every batch one by one based a short test run on their systems. This gives approximate results.
The lower the number of atoms, the utilization (~credit/time ratio) of faster GPUs will be lower.
All of the above together explains the different credits and the different credit/time ratio.
When you compare the credit/time ratio of different batches (done by the same app), you compare apples to oranges.
Look at the end of the stderr.txt output for clues:
ADRIA_FOLDADAH2
PABLO_2IDP
ADRIA_FOLDUCB

name of Granted Number of avg time number of BATCH Credits steps per step atoms ADRIA_FOLDADAH2: 181,050 12,50M 4.059ms 32,198 PABLO_2IDP: 110,400 10,00M 3.259ms 24,524 ADRIA_FOLDUCB: 63,750 12,26M 1.932ms 11,340

Maybe it's the same car,
but not the same road, not the same weather, not the same fuel, not the same driver, and not the same other cars in the road.
Hope that helps to understand this.

Erich56
Send message
Joined: 1 Jan 15
Posts: 578
Credit: 2,998,011,444
RAC: 1,277,936
Level
Phe
Scientific publications
watwatwatwatwatwat
Message 51376 - Posted: 26 Jan 2019 | 14:16:45 UTC

Something seems wrong with the credit calculation:

today, a WU did NOT get granted the "below 24 hours" extra bonus of 20%.
The total time from download till upload was 23 hours and 35 minutes. I received the 20% bonus for "below 48 hours", and that was it.
http://gpugrid.net/result.php?resultid=20416897

Okay, no big deal, but my efforts to make to make it below 24 hours, like giving it an extra GPU glock speed, were for nothing :-(

Profile Retvari Zoltan
Avatar
Send message
Joined: 20 Jan 09
Posts: 2031
Credit: 14,709,489,669
RAC: 1,510,420
Level
Trp
Scientific publications
watwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwat
Message 51378 - Posted: 26 Jan 2019 | 16:17:52 UTC - in response to Message 51376.

The total time from download till upload was 23 hours and 35 minutes.
According to the task you've linked it was 24 hours and 1 minute and 8 seconds:
Sent: 25 Jan 2019 9:59:05 UTC Received: 26 Jan 2019 10:00:13 UTC
The processing of the workunit took 84,544 seconds which is 23h 29m 04s.
I received the 20% bonus for "below 48 hours", and that was it.
http://gpugrid.net/result.php?resultid=20416897
Perhaps the download of the task and/or the upload of the result took more than normal. I see a couple of stalled uploads on my hosts lately, also the project's website can't be reached sometimes on my hosts. However if I try to reach it from a different ISP at the same time, there's no problem.

Profile Beyond
Avatar
Send message
Joined: 23 Nov 08
Posts: 1104
Credit: 6,101,732,079
RAC: 890
Level
Tyr
Scientific publications
watwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwat
Message 51379 - Posted: 26 Jan 2019 | 17:44:33 UTC - in response to Message 51376.

The total time from download till upload was 23 hours and 35 minutes. I received the 20% bonus for "below 48 hours", and that was it.
http://gpugrid.net/result.php?resultid=20416897

Okay, no big deal, but my efforts to make to make it below 24 hours, like giving it an extra GPU glock speed, were for nothing :-(

Unfortunately I ran into this literally hundreds of times when I was running 17 750ti GPUs. Due to some consideration by the Furies and other WU length Gods the 750Ti often just misses the 24 mark even when using a 0 project priority. Here's the WU time in question:

20416897 205584 25 Jan 2019 | 9:59:05 UTC 26 Jan 2019 | 10:00:13 UTC

You missed the 24hr bonus by 68 seconds. The time is calculated from when the WU download starts to when it is reported (not uploaded). Unfortunately downloads often stall numerous times even on fast connections (this has been going on forever on this project and only on this project AFAIK). Uploads take a while because they're huge. Yes, it's irritating if you're running 750Ti GPUs. I empathize with you.

tullio
Send message
Joined: 8 May 18
Posts: 157
Credit: 37,709,395
RAC: 12,901
Level
Val
Scientific publications
wat
Message 51380 - Posted: 26 Jan 2019 | 18:58:23 UTC

I have downloaded a QC task on my main Linux box. It is there, waiting to run. No other task is running. Disk space is abundant.
Tullio

Erich56
Send message
Joined: 1 Jan 15
Posts: 578
Credit: 2,998,011,444
RAC: 1,277,936
Level
Phe
Scientific publications
watwatwatwatwatwat
Message 51385 - Posted: 27 Jan 2019 | 13:55:47 UTC - in response to Message 51378.

Zoltan wrote:

Perhaps the download of the task and/or the upload of the result took more than normal. I see a couple of stalled uploads on my hosts lately, also the project's website can't be reached sometimes on my hosts.

Although I experienced upload stalls several times before (and reported about them here), I didn't notice them in the recent past.
However, I am - of course - not present each time a WU of my total 5 hosts is being uploaded, so it could well have happened.
In fact, I couldn't imagine any other reason.

Dusan
Send message
Joined: 30 Jan 19
Posts: 1
Credit: 14,802,263
RAC: 25
Level
Pro
Scientific publications
wat
Message 51485 - Posted: 14 Feb 2019 | 1:56:45 UTC
Last modified: 14 Feb 2019 | 2:17:55 UTC

Hello, I have a similar problem (or rather anomaly) - my credit got quadrupled today from ~70k to ~280k. I noticed that I finished a long GPU task, but why is this one particular task worth so much? How to find out?

Edit: found it - http://www.gpugrid.net/result.php?resultid=20499467


____________
I am a software developer and I know absolutely nothing about medicine. So I want to help those, who do know :)

CPU: AMD Ryzen 5 1600
GPU: Nvidia GTX 1050Ti (upgrade to 1070 coming soon™)

Post to thread

Message boards : Number crunching : Granted credits