PrimeGrid Challenges 2021

Page 7 - Seeking answers? Join the AnandTech community: where nearly half-a-million members share solutions and discuss the latest tech.

StefanR5R

Elite Member
Dec 10, 2016
5,459
7,718
136
I looked up the current search range of Seventeen or Bust and checked its FFT sizes on a Haswell CPU. Haswell and all later Intel CPUs which don't implement AVX-512, and several AMD CPU architectures including Zen 2 and certainly Zen 3, will use the same FFT implementation here, which is "all-complex FMA3 FFT".
I was wrong in one aspect:
While both Haswell and Zen 2 use "all-complex FMA3 FFT", they still may use different FFT lengths for the same candidate. I started tests with 24737*2^34368511+1 now, which uses 3M FFT on Broadwell-EP, but 3200K FFT on Zen 2.
 

waffleironhead

Diamond Member
Aug 10, 2005
6,917
429
136
If anyone wants a reference time, my 3700x using all 8 cores(no SMT) does sob at 2.88% per hour or about 125,000 seconds(35 hours).
 
  • Like
Reactions: crashtech

StefanR5R

Elite Member
Dec 10, 2016
5,459
7,718
136
I found in my Zen 2 llrSOB tests that temperatures of the RAM modules were ~5...10 degrees higher than when TN-Grid was running, and temperatures of the RAM VRMs were ~10...15 degrees higher.

Zen 2 will perform poorly at this, because its level-3 cache is divided into 16 MB segments. (In APUs it is much less.) Zen 2 will therefore have to perform a lot of main memory accesses and will consequently be unable to keep its FMA units busy.
After I actually tried it, I have to say it is not too bad after all, at least on my 32-core Epyc and on @waffleironhead's Ryzen which both have a ratio of 4 cores : 1 memory channel. (In my case: 4.8 W power budget per core, and DDR4-3200c22 memory.)

I guess the Zen 2 CPUs with 8 cores : 1 memory channel will show clearer signs of memory bandwidth limitation.

If anyone wants a reference time, my 3700x using all 8 cores(no SMT) does sob at 2.88% per hour or about 125,000 seconds(35 hours).
Ah, I see there are already four complete results on your host, each one apparently in the 8-threaded config,
152,931.14 s (42 h) for 74,122.76 credit -> 42,000 PPD
124,224.46 s (35 h) for 74,126.64 credit -> 52,000 PPD
125,101.35 s (35 h) for 79,209.90 credit -> 55,000 PPD
128,091.88 x (36 h) for 74,165.35 credit -> 50,000 PPD​

I tried the "8cores(no SMT)" config (among other configs) on the 7452 with a 74,066.82 credit WU but only until it was 1 % complete. Simply taking the test duration times 100, I arrive at 43 h duration and 42,000 PPD (per each set of 8 cores; 330,000 PPD over all 2x32 cores).
 

StefanR5R

Elite Member
Dec 10, 2016
5,459
7,718
136
I tried the "8cores(no SMT)" config (among other configs) on the 7452 with a 74,066.82 credit WU but only until it was 1 % complete. Simply taking the test duration times 100, I arrive at 43 h duration and 42,000 PPD (per each set of 8 cores; 330,000 PPD over all 2x32 cores).
Some more results from my 1% tests with this 74,066.82 credit WU — the following are all linear extrapolations by taking the test results times 100:
  • The dual EPYC 7452 setup for highest throughput gives an estimated 340 kPPD.
  • The dual EPYC 7452 setup for highest efficiency causes 6.9 MJ energy consumption per task.
  • On dual Xeon E5-2696 v4 with permanently enabled all-core turbo, the setup which gives best throughput is at the same time giving highest efficiency: 365 kPPD earnings, 9.5 MJ consumption per task.
 

VirtualLarry

No Lifer
Aug 25, 2001
56,229
9,990
126
Planning to jump into this tomorrow, with 2x Zen2 3600, and I've got one 3800X running a mining rig with 4x RX 5700XT cards, might put that CPU onto SoB/PrimeGrid as well.

I think that I'm going to let my 1600 sit this one out, it's getting slowly flakier.
 
  • Like
Reactions: Ken g6

StefanR5R

Elite Member
Dec 10, 2016
5,459
7,718
136
What's weird is that my computers used sllr2_1.1.0_linux64_201114 during the January challenge at PPS-DIV, but one test computer which fetched two SoB-LLR tasks today to test the waters downloaded and uses sllr2_1.0.0_linux64_200814 instead for SoB.
llrSOB is using sllr2_1.1.0_linux64_201114 now too.
 

Ken g6

Programming Moderator, Elite Member
Moderator
Dec 11, 1999
16,220
3,803
75
So, how many people's computers started an hour early? I know mine did. Stupid daylight saving time.
 
  • Like
Reactions: Markfw

Ken g6

Programming Moderator, Elite Member
Moderator
Dec 11, 1999
16,220
3,803
75
Day 1 stats:

Rank___Credits____Username
6______542624_____xii5ku
21_____149551_____Icecold
33_____79553______crashtech
35_____79295______Orange Kid
37_____77078______Ken_g6

Rank__Credits____Team
1_____1311893____SETI.Germany
2_____1226483____Czech National Team
3_____928102_____TeAm AnandTech
4_____925509_____Ural Federal University
5_____920037_____BOINC@MIXI
6_____913379_____Sicituradastra.

Wow! I suppose slower computers will come along to unseat us, but 3rd is incredible! :D
 

waffleironhead

Diamond Member
Aug 10, 2005
6,917
429
136
I started a SOB on my 3700x using SMT, so 8C16T unit. SMT made no difference, its still running at 2.88%hour.
My 2400g systems running 4C8T are 1.44%/hour.
For fun, I fired up boinc on an old A12-8800 system in the office, using all four "cores" its running .36%/hour. Never going to make the deadline at that rate. heh.
 

Markfw

Moderator Emeritus, Elite Member
May 16, 2002
25,483
14,434
136
I started primegrid on my 5950x to see how it does.
 

Markfw

Moderator Emeritus, Elite Member
May 16, 2002
25,483
14,434
136
You might try going into your preferences on the primegrid site and making sure you are set to use CPU.
That wasn't it, BUT I checked, and it wants only 17 or bust selected. I did that, and its now downloading !

0.113% in 13 minutes for 31 threads, ugly.

There is a config to use multiple threads per task ? What is it, and how many should I do ?
 
Last edited:

StefanR5R

Elite Member
Dec 10, 2016
5,459
7,718
136
@Markfw,
my suggestion is that
  • two tasks should be running in parallel,
    (Reason: 5950X has got two L3 caches, each 32 MB large, and llrSOB needs ~25 MB cache per task.)
  • each task uses 8 program threads.
    (Reason: On other processors with large enough cache, SMT or HT don't improve throughput. That is, there should be as many program threads as there are physical cores = 2x 8.)
I am very sure of the first suggestion, but not sure of the second suggestion. The Zen 3 core is larger than e.g. Zen 2's; perhaps the use of SMT would have a better effect with this one. Hard to find out without synthetic tests.

You can set both via the PrimeGrid web site:
"Your account" -> "PrimeGrid preferences" -> "Edit PrimeGrid preferences" -> fairly at the top of this long page:
  • "Max # of simultaneous PrimeGrid tasks" lets you set a limit on tasks in progress.
  • "Multi-threading: Max # of threads for each task" lets you define the amount of program threads per task.
(After setting these, scroll down to the end of the page and hit "Update preferences".)

The "Max # of simultaneous PrimeGrid tasks" setting enforces a server-side limit and therefore gets effective immediately for the requesting of new tasks. (But if the client loaded more tasks earlier already, those remain on the client; they won't be retracted or something.)

The "Multi-threading: Max # of threads for each task" setting affects tasks
  • only after the client contacted the server and thereby saw the new setting,
  • and after this, only when the client starts a new task or resumes a suspended task.
AFAIK. I still use the older app_config.xml approach instead.
 
Last edited:

StefanR5R

Elite Member
Dec 10, 2016
5,459
7,718
136
PS,
if you have another load on the computer, such as Folding@home, then it is probably best to just use half of the hardware threads for llrSOB. I.e. 8 program threads per task for two simultaneously running tasks.

Besides the "Max # of simultaneous PrimeGrid tasks" setting on the web site, which is misnamed and should rather be called "Max # of PrimeGrid tasks in progress per host", you can of course control locally how many tasks are running at once by setting the allowed number of logical CPUs in boincmgr accordingly. E.g. 50 % if you set 8 threads per task and want 2 tasks running simultaneously.
 

crashtech

Lifer
Jan 4, 2013
10,521
2,111
146
I really didn't expect to be doing so well. I have a lot of Zen2's in the mix that don't seem well optimized for these big tasks, yet the 3900X's are doing surprisingly well. I have no doubt my ranking will deteriorate as more work rolls in. Also I would imagine you are bunkering a bit... ;)
 

Icecold

Golden Member
Nov 15, 2004
1,090
1,008
146
I really didn't expect to be doing so well. I have a lot of Zen2's in the mix that don't seem well optimized for these big tasks, yet the 3900X's are doing surprisingly well. I have no doubt my ranking will deteriorate as more work rolls in. Also I would imagine you are bunkering a bit... ;)
How do you have your 3900x's configured as far as number of threads, number of tasks at a time, SMT on or off, etc.? I have a feeling I don't have my PC's configured optimally.
 

Markfw

Moderator Emeritus, Elite Member
May 16, 2002
25,483
14,434
136
OK, I have F@H running, only one of 6 machines, so I don't want to affect that. I am running 2 x 14, so we will see how it does.
 

crashtech

Lifer
Jan 4, 2013
10,521
2,111
146
How do you have your 3900x's configured as far as number of threads, number of tasks at a time, SMT on or off, etc.? I have a feeling I don't have my PC's configured optimally.
2 tasks of 6 threads each. I have no idea if it's optimal! Didn't have time to test. But the tasks take up a huge amount of L3 cache, so I think the 2 tasks part is right.
 

Icecold

Golden Member
Nov 15, 2004
1,090
1,008
146
2 tasks of 6 threads each. I have no idea if it's optimal! Didn't have time to test. But the tasks take up a huge amount of L3 cache, so I think the 2 tasks part is right.
Interesting, mine is configured the same way but I haven't returned any tasks from my 3900x's yet. They're at 77% currently. Do you have SMT turned off in the bios, or you're just setting Primegrid/BOINC to only use 12 threads? I have BOINC set to use 50% of available threads, and SMT on in the bios.
 

crashtech

Lifer
Jan 4, 2013
10,521
2,111
146
I have Process Lasso on my Windows machines, it's set to make the app use physical cores. Both my 3900X's are on Windows, unfortunately. To me it looked like I got some tasks that were very short early on, I'll have to check and see what that is about.

Edit:
8 out of 10 of my completed tasks were "main" tasks, which are short. Only 2 are long "proof" tasks.
 
Last edited:
  • Like
Reactions: Icecold

Markfw

Moderator Emeritus, Elite Member
May 16, 2002
25,483
14,434
136
OK, at 2 x 14, it keeps locking up. WCG ran ever since I built it (at stock) and 31 tasks and no problem.. If it could ever stay up, its says 14 hours each, how is that ?
 

Icecold

Golden Member
Nov 15, 2004
1,090
1,008
146
I appreciate the info @crashtech . I think the short ones are the double checker tasks. I'll keep an eye on mine, I did get a few hours late start so maybe my tasks will start rolling in. It just had me a bit alarmed that I'm barely turning in any tasks.

OK, at 2 x 14, it keeps locking up. WCG ran ever since I built it (at stock) and 31 tasks and no problem.. If it could ever stay up, its says 14 hours each, how is that ?
Can you try 2x8 ? Primegrid doesn't generally respond well to hyperthreading or SMT, so it will probably run better at 2 tasks with 8 threads each.