Advanced search

Message boards : Number crunching : Subset of Workunits Giving Lower Credit

Author Message
Scott Brown
Send message
Joined: 21 Oct 08
Posts: 144
Credit: 2,973,555
RAC: 0
Level
Ala
Scientific publications
watwatwatwatwatwat
Message 7205 - Posted: 4 Mar 2009 | 22:07:36 UTC

I had suspected that a couple of types of workunits were giving lower credit than others for some time. By this, I am not referring to the different credit awards (e.g., the 24xx vs. 37xx, etc.). Rather, some work seems to be systematically awarding less credit per hour than other types within the same general award type (i.e., some of the 24xx units award less credit per hour than others, etc.). To check this, I examined all recent work that I have completed by work type and reported runtimes. I report my findings for my 9600GSO (I confirmed the patterns as much as was possible with my two slower cards, a 9500GT and an 8800GS).

Card Specs:
ASUS EN9600GSO TOP (factory OC)
650 core, 1700 shader, 1800 memory (runs 24/7)

Workunit data from Jan. 28 - March 4, 2009
All work completed using BOINC 6.5.0 (GPURID App. 6.62)

Workunits with the following names award about 178 credits per hour: US_3, US_6, US_7, US_9, US_10, US_11
*note that these are both 24xx and 37xx credit award units
**note also that US_8 and SMDxx work on the other cards (9500GT and 8800GS) was consistent with this credit rate comparison on those cards.


Workunits with the following names award about 160 credits per hour: US_2, US_5, GRA1
*note this group includes 24xx and 32xx credit award units

On the 9600GSO, this equates to about 400-500 less credit per day. On faster cards from the GTX series, this difference would be much larger. (8800GS with 1375 shader clock has about 300 credit per day difference; 9500GT has about 200 credit per day difference).

Please note that I do not point out this difference as a complaint. Rather, since it is my understanding that the project calculates the fixed credit using a single algorithm across all types of work, these different types of work my differ in a systematic way that needs to be accounted for in the algorithm, especially since future work may have related issues.


ms113
Send message
Joined: 6 Feb 09
Posts: 19
Credit: 1,281,738
RAC: 0
Level
Ala
Scientific publications
watwatwatwatwat
Message 7212 - Posted: 5 Mar 2009 | 9:16:37 UTC - in response to Message 7205.

I can confirm this behavior.

Using Linux x86_64, Boinc 6.4.5., App.v. 6.59 and a 9800GT the different WUs awarded as follows:

US_2 WUs awarded about 200 credits per hour.
(.. using 9% of one CPU.)

US_3, US_6, US_9, US_10, US_12 WUs awarded about 225 credits per hour.
(.. using 40% of one CPU.)

Interestingly, since March 1, 2009 this changed slightly:

.. since 4 days, e.g. the US_3, US_6, US_8 are using 36% of one CPU and award now about 220 credits per hour.

I cannot tell about the actual behavior of US_2, since I am aborting them (due to your suggestion) since Feb. 26 (they feel to choppy on my system!!)

This all seems to be related to my questions in this thread concerning CPU-Utilization and choppiness.

Regards,
Martin

Post to thread

Message boards : Number crunching : Subset of Workunits Giving Lower Credit

//