Proposed Enhancement for TOP GPU MODELS Page

Questions and Answers : Wish list : Proposed Enhancement for TOP GPU MODELS Page
Message board moderation

To post messages, you must log in.

AuthorMessage
Profile David C Blanchard

Send message
Joined: 14 Sep 00
Posts: 9
Credit: 94,757,692
RAC: 0
United States
Message 1587577 - Posted: 16 Oct 2014, 3:05:18 UTC

I'm seeing a lot of day-to-day variability in the relative performance of various GPU models on the TOP GPU MODELS page which makes me suspect that I may be mis-interpreting it.

I thought (hoped) it was a performance comparison, normalized by quantity for each model type. However, the wild variations I'm seeing suggest that it is instead reporting the relative contribution, by group, of however many GPUs of a given type happen to be reporting.

PROPOSAL -- Since BOTH interpretations are valid but different, perhaps a column could be adding to indicate the number of GPUs represented (in addition to their combined contribution).

This would allow people like me to see (1) The relative popularity of each GPU model, and (2) provided the means to normalize the comparitive contributions. For example, if 10 GTX 670s get a relative rating of 0.4 and 20 GTX 660s have a relative rating of 0.6 I would know that a 670 is actually 33% more powerful than a 660--at least for SETI work!

David C Blanchard (Founder of FAA WJHTC) dblanch256@gmail.com
ID: 1587577 · Report as offensive
Profile BilBg
Volunteer tester
Avatar

Send message
Joined: 27 May 07
Posts: 3720
Credit: 9,385,827
RAC: 0
Bulgaria
Message 1587821 - Posted: 16 Oct 2014, 15:26:52 UTC - in response to Message 1587577.  
Last modified: 16 Oct 2014, 15:42:20 UTC

ID: 1587821 · Report as offensive
Profile David C Blanchard

Send message
Joined: 14 Sep 00
Posts: 9
Credit: 94,757,692
RAC: 0
United States
Message 1588141 - Posted: 17 Oct 2014, 5:09:20 UTC

BilBg -- Thanks for those links!

I think Wiggo nailed it when he said:

"The "most productive GPUs" is based on the amount of work returned by a certain model card so its really based on numbers and popularity of that card.
It does not reflect the actual performance of a card."

My point was that the addition of number GPUs reporting, by model, would allow people to normalize the results to compare performance by model (if only for SETI work). Armed with that information, and allowing for variations of platform architecture, the only other purchase consideration is cost.

-- Dave
ID: 1588141 · Report as offensive
Profile BilBg
Volunteer tester
Avatar

Send message
Joined: 27 May 07
Posts: 3720
Credit: 9,385,827
RAC: 0
Bulgaria
Message 1588244 - Posted: 17 Oct 2014, 10:46:28 UTC - in response to Message 1588141.  

... so its really based on numbers and popularity of that card.

That is not right

The person to listen about BOINC internals is Josef W. Segur

// "best-performing" is defined as minimizing the average of
//
// elapsed_time(J)/rsc_fpops_est(J)
// over completed jobs J currently in the DB


"Best GPU" = Min (Run-time/FLOP-count-estimate)

Seems to be calculated only for:
- tasks that run > 100 seconds
- tasks sent in the last 30 days
- only max 500 tasks for given GPU model are included in the average

Do not depend on "numbers and popularity of that card"
Only if really low number of some GPU model is used here (e.g. < 10) this may be the reason of more fluctuating
 


- ALF - "Find out what you don't do well ..... then don't do it!" :)
 
ID: 1588244 · Report as offensive
Josef W. Segur
Volunteer developer
Volunteer tester

Send message
Joined: 30 Oct 99
Posts: 4504
Credit: 1,414,761
RAC: 0
United States
Message 1588392 - Posted: 17 Oct 2014, 18:14:51 UTC

I'll add a few brief comments, but I don't really see a way to quickly make the list very meaningful.

The basic logic hasn't changed over time, though the current version does include changes such as showing Intel GPUs and appending the date and time the list was last updated at the bottom of the page.

1. The list is generated from GPUs running the stock app versions only. The users who have the S@H top hosts are quite likely to be running anonymous platform.

2. Any GPU which deserves to be in the top models list is more productive at this project when running more than one task per GPU. A BOINC app_config.xml file can now be used to enable that for stock applications, but those settings are AFAICT not reported back to the BOINC servers. If a top GPU were able to do tasks in 1000 seconds one at a time or 1500 seconds two at a time it would be rated higher when running one at a time (3 tasks in 3000 seconds) even though it is more productive when doing two at a time (4 tasks in 3000 seconds).

3. The rsc_fpops_est for different apps is never guaranteed to be comparable, but the top GPU models code ignores that. When Astropulse was in early testing at SETI Beta, an attempt was made to roughly match Astropulse to SETI@home Enhanced in that regard for CPU processing, but users judged it was off by approximately a factor of 2. Subsequent changes including those for SETI@home v7 and Astropulse v7 have probably made the mismatch worse, and a calibration for CPU is not likely to be any place near accurate for GPU anyhow.
                                                                  Joe
ID: 1588392 · Report as offensive
Profile David Anderson
Volunteer moderator
Project administrator
Project developer
Project scientist
Avatar

Send message
Joined: 13 Feb 99
Posts: 173
Credit: 502,653
RAC: 0
Message 1589169 - Posted: 20 Oct 2014, 5:37:47 UTC

I fixed a bug in the script, and changed it to include anonymous platform hosts. It doesn't reflect configs that run multiple jobs per GPU; currently this info is not available on the server.

The numbers in parents show relative speed, not total computing.
ID: 1589169 · Report as offensive
Profile David C Blanchard

Send message
Joined: 14 Sep 00
Posts: 9
Credit: 94,757,692
RAC: 0
United States
Message 1589375 - Posted: 20 Oct 2014, 19:01:22 UTC

Thank you ... and the others who posted. I now feel I understand the content.
ID: 1589375 · Report as offensive

Questions and Answers : Wish list : Proposed Enhancement for TOP GPU MODELS Page


 
©2024 University of California
 
SETI@home and Astropulse are funded by grants from the National Science Foundation, NASA, and donations from SETI@home volunteers. AstroPulse is funded in part by the NSF through grant AST-0307956.