Advanced search

Message boards : Number crunching : In-Progress Tasks that have gone past the deadline and daily Quota.

Author Message
archeye
Send message
Joined: 10 May 13
Posts: 4
Credit: 6,307,450
RAC: 27,854
Level
Ser
Scientific publications
wat
Message 52988 - Posted: 16 Nov 2019 | 16:59:58 UTC

My basic question is what is the correct thing to do for the project?

It seems like when a completed task is over its deadline and then submitted the validation is not possible. So in these cases is submission still useful?

I am thinking the right thing to do is to Abort when I see an in-progess task that is over the deadline. Is this correct?

The secondary point is about my Quota. From reading a post in another forum it was suggested that submitting an overdue task will result in your daily quota being reduced by 1.

Now I have just submitted 2 overdue tasks and when I only get 2 tasks at a time then what will happen?
____________

Profile Retvari Zoltan
Avatar
Send message
Joined: 20 Jan 09
Posts: 2077
Credit: 15,123,039,983
RAC: 4,743,651
Level
Trp
Scientific publications
watwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwat
Message 52989 - Posted: 16 Nov 2019 | 17:36:59 UTC - in response to Message 52988.
Last modified: 16 Nov 2019 | 17:37:32 UTC

My basic question is what is the correct thing to do for the project?
To return the result within 5 days.

It seems like when a completed task is over its deadline and then submitted the validation is not possible.
So in these cases is submission still useful?
These tasks are resent to another host. The result which returns sooner will be used.

I am thinking the right thing to do is to Abort when I see an in-progess task that is over the deadline. Is this correct?
No. You should check the progress of your tasks, and leave your host turned on if necessary to complete the task within the deadline.
If it bothers you, then you should choose another project. There's no point to process tasks for 15 hours in vain.

The secondary point is about my Quota. From reading a post in another forum it was suggested that submitting an overdue task will result in your daily quota being reduced by 1.
Yes, but your quota is incremented by 1 every day.

Now I have just submitted 2 overdue tasks and when I only get 2 tasks at a time then what will happen?
You can check the quota of your host for each app type on its status page.
The present quota ("Max tasks per day") is 30 for "Long runs (8-12 hours on fastest card) 9.23 windows_intelx86 (cuda80)".
The scheduler sends only 2 tasks per GPU for each host.

BTW, one of your overdue task has been assigned to one of my hosts:
http://www.gpugrid.net/workunit.php?wuid=16849143
It returned the result 12 hours earlier than yours.

Profile ServicEnginIC
Avatar
Send message
Joined: 24 Sep 10
Posts: 78
Credit: 1,248,477,976
RAC: 743,892
Level
Met
Scientific publications
watwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwat
Message 52990 - Posted: 16 Nov 2019 | 19:59:06 UTC
Last modified: 16 Nov 2019 | 20:01:14 UTC

It seems like when a completed task is over its deadline and then submitted the validation is not possible.
So in these cases is submission still useful?

These tasks are resent to another host. The result which returns sooner will be used.

By any chance, I catched a few days ago an example that overseeds both deadline and bonus criteriums at the same time.
Take a careful look to the following Work unit.
I send its link in image format, because data is erased from server database after a week past.
http://www.servicenginic.com/Boinc/GPUGrid/Forum/WU_01.png

- Firstly, this Work unit was sent to computer ID 513534 on 29/10/2019 - 12:18:35 UTC
- Deadline was arrived with no response on 03/11/2019 - 12:18:35 UTC
- Then it was resent to computer ID 105442 (one of mine) on 03/11/2019 - 12:36:53 UTC
- The first computer returned its result on 03/11/2019 - 18:34:28 UTC. This is more than 6 hours past deadline. And it was granted with standard credit amount 194.400
- The second computer returned its result on 04/11/2019 - 04/11/2019 - 04:15:48 UTC. In time for +50% credit bonus, but it was granted with standard credit amount 194.400 (instead of 291.600 with bonus added)

My conclusions:
-1) A Task that overdues deadline won't receive credits unless there is no valid result at the moment of reporting it.
-2) A task will be granted with credit bonus if it meet the criteriums for this, unless same Work unit be reported without bonus first.
-3) As Retvari Zoltan said, "canonical result" will be taken from first task reported.

Profile Retvari Zoltan
Avatar
Send message
Joined: 20 Jan 09
Posts: 2077
Credit: 15,123,039,983
RAC: 4,743,651
Level
Trp
Scientific publications
watwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwat
Message 52991 - Posted: 16 Nov 2019 | 21:21:32 UTC - in response to Message 52990.

Your conclusions are correct.
This is a known side effect of how work is treated by the scheduler, and the bonus system of GPUGrid.
A given workunit can't have two different credits awarded, hence when the host in overdue reports the result while the second host is processing the same workunit (so the workunit is still active), the credits without bonus get assigned to the workunit. When the second host reports the result, the credits awarded won't be calculated again.
If the first host reports the result of the task in overdue while it waits in the queue of the second host, then the task will be cancelled on the second host (provided that it connects to the project before it starts that workunit).

archeye
Send message
Joined: 10 May 13
Posts: 4
Credit: 6,307,450
RAC: 27,854
Level
Ser
Scientific publications
wat
Message 52992 - Posted: 17 Nov 2019 | 8:19:59 UTC - in response to Message 52989.
Last modified: 17 Nov 2019 | 8:23:47 UTC

Thank you Retvari for you complete and informative reply it was very useful.

I crunch the tasks while I have the computer on during the day and the only reason these 2 tasks did not complete within the time was I needed to be away for 2 days after they started and the computer remained off.

The scheduler sends only 2 tasks per GPU for each host.


I have 2x nVidia GTX 980 GPU cards so your comment suggests I should be getting 4 tasks at a time however I am ok with just 2 tasks at a time because with the PC ontime I allow my GPUs would not be able to complete all 4 within the timeframe given.

The PC is in the next room to where I sleep and the high pitched whine is just too much for me at night so I will not keep anything running overnight.

This is now off topic of the post but when cuda80 is referenced for tasks how do I find the cuda number for my GTX 980's and what does this tell me performance wise for crunching?

Profile Retvari Zoltan
Avatar
Send message
Joined: 20 Jan 09
Posts: 2077
Credit: 15,123,039,983
RAC: 4,743,651
Level
Trp
Scientific publications
watwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwat
Message 52993 - Posted: 17 Nov 2019 | 10:39:38 UTC - in response to Message 52992.

The scheduler sends only 2 tasks per GPU for each host.

I have 2x nVidia GTX 980 GPU cards so your comment suggests I should be getting 4 tasks at a time
There's not enough work on the server at the moment. This is also limited by the settings of your profile and the local settings of BOINC manager (the latter takes priority). The other limiting factor is that the BOINC manager estimates the processing time of a workunit, and requests only that many tasks that fill the remaining time of its queue.

This is now off topic of the post but when cuda80 is referenced for tasks how do I find the cuda number for my GTX 980's and what does this tell me performance wise for crunching?
You can find your CUDA (version btw) number in the first lines of the event log of BOINC manager at its startup.
2019. 11. 17. 10:38:00 | CUDA: NVIDIA GPU 0: GeForce GTX 950M (driver version 441.12, CUDA version 10.2, compute capability 5.0, 2048MB, 1680MB available, 1188 GFLOPS peak) 2019. 11. 17. 10:38:00 | OpenCL: NVIDIA GPU 0: GeForce GTX 950M (driver version 441.12, device version OpenCL 1.2 CUDA, 2048MB, 1680MB available, 1188 GFLOPS peak)
CUDA80 (=8.0) is the minimum requirement for these tasks to run, higher CUDA number of the driver doesn't mean better performance, as the present app uses the feature set of CUDA80 only (which is included in later versions). When the app will be updated to CUDA 10.1, then (in theory) it can utilize the (performance) advantage of CUDA 10.1, provided that the card's hardware can support it. This is determinded by the "Compute Capability" number of the card.

archeye
Send message
Joined: 10 May 13
Posts: 4
Credit: 6,307,450
RAC: 27,854
Level
Ser
Scientific publications
wat
Message 53004 - Posted: 20 Nov 2019 | 3:45:00 UTC - in response to Message 52993.

Well I never noticed this info hidden in the messages before, thanks :)

Now I see my GPU's are 10.1 so they will be ok if and when you update the app as long as I guess the compute capability of 5.2 is ok as well.

    PC

    7 20/11/2019 04:33:12 CUDA: NVIDIA GPU 0: GeForce GTX 980 (driver version 436.48, CUDA version 10.1, compute capability 5.2, 4096MB, 3292MB available, 4979 GFLOPS peak)
    8 20/11/2019 04:33:12 CUDA: NVIDIA GPU 1: GeForce GTX 980 (driver version 436.48, CUDA version 10.1, compute capability 5.2, 4096MB, 3292MB available, 4979 GFLOPS peak)
    9 20/11/2019 04:33:12 OpenCL: NVIDIA GPU 0: GeForce GTX 980 (driver version 436.48, device version OpenCL 1.2 CUDA, 4096MB, 3292MB available, 4979 GFLOPS peak)
    10 20/11/2019 04:33:12 OpenCL: NVIDIA GPU 1: GeForce GTX 980 (driver version 436.48, device version OpenCL 1.2 CUDA, 4096MB, 3292MB available, 4979 GFLOPS peak)

Post to thread

Message boards : Number crunching : In-Progress Tasks that have gone past the deadline and daily Quota.