Different workunit sizes added

log in

Advanced search

Message boards : News : Different workunit sizes added

1 · 2 · Next
Author Message
Teemu Mannermaa
Project administrator
Project developer
Project tester
Send message
Joined: 20 Apr 11
Posts: 355
Credit: 718,947,145
RAC: 177,932
Message 423 - Posted: 22 May 2011, 12:39:09 UTC

Hi,

I just completed adding different workunit sizes and scheduler should now sent you work based on the measured speed of your host. Additionally, scheduler sents workunits that are better "match" for number of cards your host has. This should minimize idle cards at the end of wu.

I'm pretty sure there's still some tweaking to do and I will be watching how the scheduler performs tomorrow. Please, do let me know if there seems to be something odd with workunits given to your host and especially if you are now unable to get any work. Thanks!

-w

Senilix
Send message
Joined: 11 May 11
Posts: 26
Credit: 50,020,469
RAC: 0
Message 424 - Posted: 22 May 2011, 13:12:54 UTC - in response to Message 423.
Last modified: 22 May 2011, 13:15:05 UTC

I don't know if this is related to your latest change but i got a download error on one of the new wus (all others downloaded successfully)

22.05.2011 15:07:26 | Moo! Wrapper | [error] MD5 check failed for dnetc_r72_1306054748_9_576
22.05.2011 15:07:26 | Moo! Wrapper | [error] expected c756e5606cafbfa8a33456ef524199d1, got b4a20620e391bd3ba1b3a4b105f4c631
22.05.2011 15:07:26 | Moo! Wrapper | [error] Checksum or signature error for dnetc_r72_1306054748_9_576

Edit: i just noticed that the donwload failed for my wingman too.

Thamir Ghaslan
Send message
Joined: 21 May 11
Posts: 3
Credit: 38,762,992
RAC: 0
Message 425 - Posted: 22 May 2011, 14:17:09 UTC - in response to Message 423.

Hi,

I just completed adding different workunit sizes and scheduler should now sent you work based on the measured speed of your host. Additionally, scheduler sents workunits that are better "match" for number of cards your host has. This should minimize idle cards at the end of wu.

I'm pretty sure there's still some tweaking to do and I will be watching how the scheduler performs tomorrow. Please, do let me know if there seems to be something odd with workunits given to your host and especially if you are now unable to get any work. Thanks!

-w


Thanks, its been smooth sailing ever since I attached yesterday.

One tiny bit of observation, sending an even number of packets in each WU might benefit both single and dual GPU setups, might not for tri, and might for quads. Provided all cards in multi GPU are of same speed.

For instance, I got some WU with 3 or 9 packets, and one of my 5970 GPU idles with these kind of WUs.

Bryan @ SUSA
Send message
Joined: 2 May 11
Posts: 15
Credit: 341,851,123
RAC: 0
Message 427 - Posted: 22 May 2011, 15:18:21 UTC - in response to Message 423.

Hi,

This should minimize idle cards at the end of wu.


We'll see how well it works, but regardless, I for one appreciate the attempt at maximizing our hw!

Thank you

Teemu Mannermaa
Project administrator
Project developer
Project tester
Send message
Joined: 20 Apr 11
Posts: 355
Credit: 718,947,145
RAC: 177,932
Message 430 - Posted: 22 May 2011, 21:30:09 UTC - in response to Message 424.

I don't know if this is related to your latest change but i got a download error on one of the new wus (all others downloaded successfully)


During my changes there were some workunits generated that have modified input files. These will fully error out after five failures so they should not bother us forever. Apologies for these!

-w

Profile [AF>EDLS] Polynesia
Avatar
Send message
Joined: 1 May 11
Posts: 23
Credit: 1,574,433
RAC: 0
Message 431 - Posted: 22 May 2011, 23:14:57 UTC

It is not possible to select the unit size?
____________

Config : i7 860 2.8ghz, 8g ram, boinc : 6.12.26, GPU : GTX 470 Zotac Amp Edition 1280 mo DDR5

Teemu Mannermaa
Project administrator
Project developer
Project tester
Send message
Joined: 20 Apr 11
Posts: 355
Credit: 718,947,145
RAC: 177,932
Message 433 - Posted: 23 May 2011, 6:49:50 UTC - in response to Message 431.

It is not possible to select the unit size?


No, not at the moment but everybody should get units that are around 30min to max few hours long (this is the intention anyway).

Could you tell me why you would want to select the size? Are there many others who want this feature?

-w

Profile [AF>EDLS] Polynesia
Avatar
Send message
Joined: 1 May 11
Posts: 23
Credit: 1,574,433
RAC: 0
Message 435 - Posted: 23 May 2011, 8:25:02 UTC - in response to Message 433.

It was so one day I want to calculate a larger unit than an hour as before ....
____________

Config : i7 860 2.8ghz, 8g ram, boinc : 6.12.26, GPU : GTX 470 Zotac Amp Edition 1280 mo DDR5

Profile ZydorProject donor
Avatar
Send message
Joined: 5 May 11
Posts: 233
Credit: 351,414,150
RAC: 0
Message 440 - Posted: 23 May 2011, 11:34:22 UTC - in response to Message 423.

..... Please, do let me know if there seems to be something odd with workunits given to your host .....


I had a 2 day cache set for my second machine with 1 x 5850 in it (should have been a 1 day cache - oppps sorry, I've changed it now, its been happily crunching away for days without a hickup I hadnt noticed), and started getting downloads of the new ones this morning. Got a bunch of six that failed to download at 8.41 UTC.

The others between then and now appeared to come in ok, so dont know if it was a one off my end, or something at the server ... anyway, for what its worth (or not as the case may be!), they are shown in the error section for that machine. All else appears fine, although I have not yet got to the new ones to crunch on that machine, that will not be for a day or so until I reduce that silly cache .....

My main machine appears to be happy enough, although I still get way inflated timing of completion recorded for the 2x5970s, even though the reality is much faster completion. Recorded times appear to be almost double the reality time to crunch (the overall time inside the result file is correct, its the time recorded in the Valid GPU time column thats way out. It could be because I am running crossfire - which for me works better (except the spurious timings). That recorded time in Valids doesnt affect me, its just a curiosity, everything else is fine - stats etc are ok.

Regards
Zy

Profile Beyond
Avatar
Send message
Joined: 18 May 11
Posts: 46
Credit: 1,250,865,589
RAC: 262,738
Message 442 - Posted: 23 May 2011, 13:06:45 UTC

I'd really like to see an option to run 1 WU/GPU. Other than that things are working well.

[BOINCstats] LostBoy
Send message
Joined: 7 May 11
Posts: 2
Credit: 5,015,275
RAC: 0
Message 443 - Posted: 23 May 2011, 14:12:02 UTC

Hi
since the new option,i am getting 98% of the short _192 ones.
They are running only about 15min on my overclocked 5750.
I was getting the _6xx before this, which ran about one hour and this was fine for me.
So an option in the preferences to select the lenght would be really fine.

Or is it possible to write an app_info.xml to adjust the flops by myself to get the long running ones,like i can do for primegrid to adjust the cpu/gpu runtime estimation ratio?

Thank you
Michael
____________

Copycat-Digital for WCG*Project donor
Avatar
Send message
Joined: 11 May 11
Posts: 44
Credit: 291,412,341
RAC: 0
Message 450 - Posted: 23 May 2011, 19:11:54 UTC - in response to Message 443.


They are running only about 15min on my overclocked 5750.


Same here. Pages of _192 WUs on a rig with an OCd 5770. Only 12 min

The longer ones ( _576 up )is nice for the 58xx's

Profile ZydorProject donor
Avatar
Send message
Joined: 5 May 11
Posts: 233
Credit: 351,414,150
RAC: 0
Message 451 - Posted: 23 May 2011, 20:48:19 UTC
Last modified: 23 May 2011, 20:56:53 UTC

Just had one stick for 40 mins on 2x5970s before I picked it up, dont yet know if its isolated or if more will follow

http://moowrap.net/result.php?resultid=288561

Edit:
Next went through fine (big one 768 stat) that took 460 secs but reported time in GPU column was 30 mins. Another trundling through fine (big 768 stat), looks like a one off fail for some reason.

Regards
Zy

Bryan @ SUSA
Send message
Joined: 2 May 11
Posts: 15
Credit: 341,851,123
RAC: 0
Message 455 - Posted: 24 May 2011, 0:56:53 UTC

I'm not having any trouble since the very start of the new format. Originally one of the 5970s was getting nothing but 3 packet wu that absolutely kill throughput :) I detached/reattached and now that machine is getting even packet wu and the large ones.

All is fine at this time.

I really appreciate the wu optimizaation. On DNETC we kept asking for 1 wu per GPU because our GPUs were sitting idle but we were IGNORED. At least now machines with 1, 2, or 4 GPUs can have everything running full time!

THANK YOU!

Alyx
Send message
Joined: 15 May 11
Posts: 3
Credit: 10,445,835
RAC: 0
Message 459 - Posted: 24 May 2011, 7:45:48 UTC

How is credit granted now? still static?

Teemu Mannermaa
Project administrator
Project developer
Project tester
Send message
Joined: 20 Apr 11
Posts: 355
Credit: 718,947,145
RAC: 177,932
Message 461 - Posted: 24 May 2011, 8:17:42 UTC - in response to Message 459.

How is credit granted now? still static?


Yes, credit is 8.5cr per stat units.

-w

Teemu Mannermaa
Project administrator
Project developer
Project tester
Send message
Joined: 20 Apr 11
Posts: 355
Credit: 718,947,145
RAC: 177,932
Message 462 - Posted: 24 May 2011, 8:39:51 UTC - in response to Message 443.

since the new option,i am getting 98% of the short _192 ones.


Looks like scheduler estimates 536.14 GFLOPS for your host (based on current completion time averages). This means you get the normal work units, because limit is 1000 GFLOPS for the huge ones.

I'm thinking about lowering that so more people will get the huge work. At the moment I was thinking 750 GFLOPS but that would mean you would still get the normal ones.. Maybe 500 GFLOPS instead then unless scheduler changes it's mind about your speed.

-w

Teemu Mannermaa
Project administrator
Project developer
Project tester
Send message
Joined: 20 Apr 11
Posts: 355
Credit: 718,947,145
RAC: 177,932
Message 463 - Posted: 24 May 2011, 8:48:08 UTC - in response to Message 425.

One tiny bit of observation, sending an even number of packets in each WU might benefit both single and dual GPU setups, might not for tri, and might for quads. Provided all cards in multi GPU are of same speed.

For instance, I got some WU with 3 or 9 packets, and one of my 5970 GPU idles with these kind of WUs.


If you have a host with only one 5970 (which means it shows up as 2 x 5870 system) you should be getting work with even number of packets. That way there shouldn't be any idle due to packets. (But can be if speed or packet size differences.)

This change should also benefit triple card systems (I have one). If they got a work with 7 packets, then it would need three cycles but with two cards idling at the last round, which is not nice. However, if they got work with 9 packets, then the system would still need three cycles to go through it but with no idle cards at the last round.

-w

Copycat-Digital for WCG*Project donor
Avatar
Send message
Joined: 11 May 11
Posts: 44
Credit: 291,412,341
RAC: 0
Message 465 - Posted: 24 May 2011, 12:03:04 UTC - in response to Message 462.


I'm thinking about lowering that so more people will get the huge work.
-w


Thanx Teemu
Getting big ones now on the HD5770

[BOINCstats] LostBoy
Send message
Joined: 7 May 11
Posts: 2
Credit: 5,015,275
RAC: 0
Message 466 - Posted: 24 May 2011, 14:29:43 UTC - in response to Message 465.


I'm thinking about lowering that so more people will get the huge work.
-w


Thanx Teemu
Getting big ones now on the HD5770


Yes,thank you very much.
I am getting the big ones now,too.

Michael
____________

1 · 2 · Next

Message boards : News : Different workunit sizes added


Main page · Your account · Message boards


Copyright © 2011-2017 Moo! Wrapper Project