Advanced search

Message boards : Number crunching : Filtering GPUs for Long Runs

Author Message
5pot
Send message
Joined: 8 Mar 12
Posts: 411
Credit: 2,083,882,218
RAC: 0
Level
Phe
Scientific publications
watwatwatwatwatwatwatwatwatwatwatwatwatwat
Message 24356 - Posted: 10 Apr 2012 | 18:29:43 UTC

Don't know if this has been brought up before, and I know everyone should be allowed to pick what they want, but seeing things like this makes me sad...http://www.gpugrid.net/workunit.php?wuid=3327127

Has GPUgrid ever thought of making long runs opt-in w/ a warning for cards on accounts page, or someway to prevent issues like the one from above from happening? (besides prevent x amount of WU from being d/l after failures.)

Profile Retvari Zoltan
Avatar
Send message
Joined: 20 Jan 09
Posts: 2343
Credit: 16,201,255,749
RAC: 7,520
Level
Trp
Scientific publications
watwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwat
Message 24384 - Posted: 11 Apr 2012 | 16:17:25 UTC - in response to Message 24356.
Last modified: 11 Apr 2012 | 16:24:42 UTC

I think you meant this.
This is not the only user, who wants to filter out NATHAN_CB1s from the long queue.
That's why I suggested crediting directly proportional only with the running time of a wu (plus 24h and 48h bounses).
Also, there should be a blacklist of hosts (with low RAC, or inadequate hardware, or selective task abortions) for long workunits.
Long workunits should be like NATHAN_FAX4, IBUCH_adTRYP, PAOLA_PYU and PAOLA_HGA.
The NATHAN_CB1 is really out of this league, though its crediting is very high.

5pot
Send message
Joined: 8 Mar 12
Posts: 411
Credit: 2,083,882,218
RAC: 0
Level
Phe
Scientific publications
watwatwatwatwatwatwatwatwatwatwatwatwatwat
Message 24385 - Posted: 11 Apr 2012 | 18:46:00 UTC

All I can really say is wow, amazing and yes I also agree CB1's don't belong in the long task feed. Again, wow

Profile Retvari Zoltan
Avatar
Send message
Joined: 20 Jan 09
Posts: 2343
Credit: 16,201,255,749
RAC: 7,520
Level
Trp
Scientific publications
watwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwat
Message 24388 - Posted: 11 Apr 2012 | 19:37:01 UTC - in response to Message 24385.

Here is another two examples of selective workunit abortions:
Host 113154
Host 113196

5pot
Send message
Joined: 8 Mar 12
Posts: 411
Credit: 2,083,882,218
RAC: 0
Level
Phe
Scientific publications
watwatwatwatwatwatwatwatwatwatwatwatwatwat
Message 24389 - Posted: 11 Apr 2012 | 20:56:49 UTC

So there are people who seriously "camp" and wait for anything but CB1's to come through, simply because they offer the most credit for the least amount of work. These CB1 should really by sent to the short task queue. That's ridiculous.

Basically, if these tasks were shifted away from long task (higher credit), than this problem would disappear? If so, I'm rather surprised this hasn't happened yet. I know nothing about scoring coding etc., but wouldn't that just be sending them to the other feeder?

Profile Retvari Zoltan
Avatar
Send message
Joined: 20 Jan 09
Posts: 2343
Credit: 16,201,255,749
RAC: 7,520
Level
Trp
Scientific publications
watwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwat
Message 24408 - Posted: 12 Apr 2012 | 15:37:38 UTC - in response to Message 24389.

So there are people who seriously "camp" and wait for anything but CB1's to come through, simply because they offer the most credit for the least amount of work.

It's strange, isn't it? These people are commonly called as credit whores. (not my words, sorry for that)

These CB1 should really by sent to the short task queue.

Personally I don't care if a series of workunit is sent to the long queue while they are not qualified to be in this queue (i.e. the project personnel abusing the long queue presumably to achieve fast result returns), unless there is another automatism which prevents abusing the long queue on the cruncher's part.

That's ridiculous.

Kind of.

Basically, if these tasks were shifted away from long task (higher credit), than this problem would disappear?

I think it would. But I would prefer if the task scheduler would qualify the host computers for processing long tasks. (beside the setting in the user profile)

If so, I'm rather surprised this hasn't happened yet.

As I mentioned above, the long queue has shorter turnaround times, so if the project need something to be finished faster, they put it in the long queue. Or, it's simply a mistake.

I know nothing about scoring coding etc., but wouldn't that just be sending them to the other feeder?

Exactly.

5pot
Send message
Joined: 8 Mar 12
Posts: 411
Credit: 2,083,882,218
RAC: 0
Level
Phe
Scientific publications
watwatwatwatwatwatwatwatwatwatwatwatwatwat
Message 24422 - Posted: 13 Apr 2012 | 3:52:42 UTC

Device 0: "NVS 300"

This one failed on a different device (NOT ONE OF MINE), attempting a FAX4. I mean I appreciate the fact that they're trying to contribute, but I actually had to look up to see what this even was. A total of 16 CUDA processors and 2 SM. I did find it interesting that it costs $130 though.

Post to thread

Message boards : Number crunching : Filtering GPUs for Long Runs

//