PrimeGrid
Please visit donation page to help the project cover running costs for this month

Toggle Menu

Join PrimeGrid

Returning Participants

Community

Leader Boards

Results

Other

drummers-lowrise
1) Message boards : Number crunching : WW & 3080 ranking? (Message 150296)
Posted 139 days ago by Profile Tuna Ertemalp
I discussed this with Michael Goetz privately. Here are the relevant excerpts, with his permission, to explain what I am seeing and to close this thread:

Michael Goetz wrote:
. . . there's a lot of things that can mess up the statistics. I think in this particular case, the problem can be boiled down to "too much of a good thing" (with "good thing" being defined as "lots of results").

The database query needs to run fairly quickly. It can't be slowing everything down, and, practically, there's no limit on how many results could theoretically be eligible to be included in these statistics. These results are cached, so only one page per day actually calculates the data, but if you're the person looking at the page when it has to recalculate the display you don't want to wait five minutes for the page to load.

Here's a query that's derived from the code in that page...

select count(*) from result where appid=33 and app_version_id in (select id from app_version where appid=33 and plan_class not like '%cpu%' and plan_class not like '%AVX%') and elapsed_time>10 and validate_state=1; +----------+ | count(*) | +----------+ | 56016 | +----------+


There's 56 thousand results that should be considered for these performance statistics. However, for performance reasons, I include a "LIMIT 3000" clause in the real query, without an "ORDER BY" clause to avoid having to sort all the rows. The database is therefore returning, in this case, a relatively small *and completely arbitrary* subset of the results.

Note that the WW challenge cleanup ended yesterday, and the WW purging was enabled, so before yesterday it wasn't 3K out of 56K results; it was 3K out of about 500K results.

Idiosyncrasies about which GPUs are included are, therefore, probably due to the small and arbitrary subset of results being chosen at a particular time when you're looking at a subproject where there's a lot of results. The timings are probably good, but you may get days where some GPUs are excluded because of bad luck.


Tuna Ertemalp wrote:
. . . Crystal-clear now. Although, my database internals work 30+ years ago makes me think that, depending on the implementation of the engine, LIMIT might keep returning arbitrary yet no-so-random data, i.e. rather stable over time. That might be one of the reasons why stuff like Ellesmere GPUs still show up, day after day.


Michael Goetz wrote:
Probably. . . . I always knew the performance numbers would be wacky. There's just too many things that will mess them up, including results that say they ran on GPU X when they actually ran on GPU Y. So crappy numbers are par for the course.

But I wasn't really expecting GPUs to go AWOL like this. Thanks for bringing it to my attention.
2) Message boards : Number crunching : WW & 3080 ranking? (Message 150261)
Posted 141 days ago by Profile Tuna Ertemalp
I believe there need to be some minimum number of cards of a type before the type shows up.

Thanks!!

Hmmm.... Assuming that's the case, methinks including the number of cards in the PG system next to each listing in parentheses would be nice. That would also make it possible to show the cards with low counts, but maybe with an (*) to indicate visually that the perf value is statistically not yet mature, while still giving the audience an idea.

Just a thought...

Tuna


I entered this on BOINC GitHub as https://github.com/BOINC/boinc/issues/4357. Seems PrimeGrid page might be having a problem. The question from BOINC dev was:

Looks like this statistics is build on last 2k results.
@TunaErtemalp, is it possible that results from 3080 are not in the last 2k batch on results of this particular app for Windows platform?


Unless my 3080s' WW results (and currently they are working on AP27, with 1000+ results so far, similarly missing from the AP27 section of that page), I'm wondering if the window of data that page is based on or frequency of its generation is having a problem...
3) Message boards : Number crunching : Suggestion for better "Top GPU Models" page (Message 150260)
Posted 141 days ago by Profile Tuna Ertemalp
As this code is part of BOINC: maybe ask the same at https://github.com/BOINC/boinc/issues?

Thanks for the link! Did so in https://github.com/BOINC/boinc/issues/4353, and seems a change is already committed, awaiting finalization of code reviews & merging into master, according to https://github.com/BOINC/boinc/pull/4354.
4) Message boards : Number crunching : Suggestion for better "Top GPU Models" page (Message 150065)
Posted 160 days ago by Profile Tuna Ertemalp
Splitting this off of the thread http://www.primegrid.com/forum_thread.php?id=9658&nowrap=true#150042.

Why are the lists split into "geforce xyz" and "nvidia geforce xyz"?

(1.000) NVIDIA GeForce RTX 3090
(0.921) GeForce RTX 3090
(0.871) GeForce RTX 3080
(0.867) NVIDIA GeForce RTX 3080


probably because this is how the driver reports the GPU name.


I wonder... On Windows, the driver comes from NVIDIA. And, all current NVIDIA drivers refer to the cards as "NVIDIA blahblah". Yet, the "NVIDIA Geforce" vs "Geforce" name duplication in the Top GPU Models page exists under the Windows columns, therefore that is not due to Linux drivers adding to the mix. Maybe the NVIDIA drivers changed sometime in the recent past to add the "NVIDIA" prefix.

In any case, the way it is currently itemized feels... ...not meaningful.

Maybe the code that generates that data that results in this page should ignore any "NVIDIA " (and, similarly "AMD ") prefix in the card name (along with the trailing space!), which would then combine the perf numbers of cards more intuitively? Separating perf numbers along an arbitrary line of when a company decided to include the company name in the driver report seems non-scientific.

Unless there is a whole another reasoning behind all of this.

Tuna

5) Message boards : Number crunching : WW & 3080 ranking? (Message 150042)
Posted 160 days ago by Profile Tuna Ertemalp
Why are the lists split into "geforce xyz" and "nvidia geforce xyz"?

(1.000) NVIDIA GeForce RTX 3090
(0.921) GeForce RTX 3090
(0.871) GeForce RTX 3080
(0.867) NVIDIA GeForce RTX 3080


probably because this is how the driver reports the GPU name.


I wonder... On Windows, the driver comes from NVIDIA. And, all current NVIDIA drivers refer to the cards as "NVIDIA blahblah". Yet, the "NVIDIA Geforce" vs "Geforce" name duplication in the Top GPU Models page exists under the Windows columns, therefore that is not due to Linux drivers adding to the mix. Maybe the NVIDIA drivers changed sometime in the recent past to add the "NVIDIA" prefix.

In any case, the way it is currently itemized feels... ...not meaningful.

Maybe the code that generates that data that results in this page should ignore any "NVIDIA " (and, similarly "AMD ") prefix in the card name (along with the trailing space!), which would then combine the perf numbers of cards more intuitively? Separating perf numbers along an arbitrary line of when a company decided to include the company name in the driver report seems non-scientific.

Unless there is a whole another reasoning behind all of this.

Tuna
6) Message boards : Number crunching : WW & 3080 ranking? (Message 150004)
Posted 162 days ago by Profile Tuna Ertemalp
What is the update frequency of the Top GPU Models page?

And, 3080 on Windows for WW showed up just this morning... Cool. Right under 3090.
7) Message boards : Number crunching : WW & 3080 ranking? (Message 149972)
Posted 162 days ago by Profile Tuna Ertemalp
Thanks!!

Hmmm.... Assuming that's the case, methinks including the number of cards in the PG system next to each listing in parentheses would be nice. That would also make it possible to show the cards with low counts, but maybe with an (*) to indicate visually that the perf value is statistically not yet mature, while still giving the audience an idea.

Just a thought...

Tuna
8) Message boards : Number crunching : WW & 3080 ranking? (Message 149970)
Posted 162 days ago by Profile Tuna Ertemalp
I have one host with dual 3080s that has generated a bunch of valid WW jobs (~400 as of this writing, over the last ~3 days): http://www.primegrid.com/results.php?hostid=782330&offset=0&show_names=0&state=4&appid=33

But http://www.primegrid.com/gpu_list.php#WW shows nothing for 3080 under Windows right now (there is only the Linux entry).

What is the update frequency of the Top GPU Models page?

Thanks
Tuna
9) Message boards : Number crunching : How to peg my newly upgraded boxes at max? (Message 149690)
Posted 177 days ago by Profile Tuna Ertemalp
as long as you are running one task per core.


Thank you!

So, you are recommending that I turn off HyperThreading on the CPUs? And, limit those projects to using just one thread per task by setting "Multi-threading: Max # of threads for each task (Only applies to LLR, AP27, and WW CPU tasks)" to "1"?

And, with "PPS" you mean just the PPS, or the category including PPS-DIV and PPS-Sieve, too?

And, with "The new WW project", are you referring to "Wieferich and Wall-Sun-Sun"?

Thanks!
10) Message boards : Number crunching : How to peg my newly upgraded boxes at max? (Message 149687)
Posted 177 days ago by Profile Tuna Ertemalp
I had been doing BOINC & PrimeGrid for a long time (1999 & 2015, respectively), but switched to FoldingAtHome for the last year, for obvious reasons. During the same year, I did upgrade the PSUs and GPUs in most of by boxes, so now I have the following. I am interested in switching back to PrimeGrid for a while to see what they are capable of. For that, I am wondering what projects I should put into my "PrimeGrid Preferences" to peg all my CPUs and GPUs at darn near 100%. During my absence from PrimeGrid, all projects progressed, and possibly new ones got added, so I am not sure how their computational needs changed. Anybody can give me the top few most CPU+GPU power hungry projects that will supply a constant stream of WUs? I am much less interested in maxing my BOINC credits or having better chances of finding primes per unit of time; I am more interested in maxing the use of my compute capacity and power supplies consistently over days/weeks, as sort of stress-test of my entire system, from cards to PSUs to breakers to incoming power service to A/C. Thanks in advance!

Here is the hardware...

OS:Win10
GPUs: EVGA

CPU (cores), RAM, GPU(s), Motherboard:

  • AMD Ryzen 5 3600 (6C), 32G DDR4-2400, 3090 FTW3 ULTRA, Gigabyte AB350M-D3H-CF
  • Intel Core i7 5960X (8C), 32G DDR4-2400, 3090 XC3 ULTRA HYBRID, ASUS X99-M WS
  • Intel Core i7 5960X (8C), 32G DDR4-2400, 2x 3090 FTW3 ULTRA, ASUS X99-E WS/USB 3.1
  • Intel Core i7 970 (6C), 24G DDR3-1333, 2x 3080 FTW3 ULTRA HYBRID, ASUS RAMPAGE III GENE
  • Intel Core i7 5960X (8C), 16G DDR4-2400, 1080 Ti FTW3 + HYBRID KIT, ASRock X99 OC Formula/3.1
  • Intel Core i7 2600 (4C), 16G DDR3-1600, 3090 FTW3 ULTRA HYBRID, ASUS P8P67
  • AMD Ryzen TR 1950X (16C), 32G DDR4-2133, 3x 3090 XC3 ULTRA HYBRID, ASRock X399 Taichi
  • Intel Core i7 5960X (8C), 64G DDR4-2133, 1080 Ti FTW3 + HYBRID KIT, 2x 1080 Ti SC2 HYBRID, MSI X99A XPOWER AC



Next 10 posts
[Return to PrimeGrid main page]
DNS Powered by DNSEXIT.COM
Copyright © 2005 - 2021 Rytis Slatkevičius (contact) and PrimeGrid community. Server load 2.85, 3.25, 3.47
Generated 20 Sep 2021 | 17:59:23 UTC