PrimeGrid Races 2018

[zzuupp]

This thread kind of caught my interest so I went ahead and dl'd BOINC and chose Primegrid from the drop down. At primegrid.com I searched for Anandtech and then joined what I believe to be the right team. How do I know though?

EDIT:I should add that I am a complete noob to this so a lot of this thread went over my head.

Welcome to the TeAm!

Ken gave you all of the details.
So just howdy and welcome again.

 

[bbhaag]

Welcome to TeAm AnandTech, bbhaag! I see you here, so you're on the right team.

I also see you've done some PrimeGrid work. However, that's not the kind of work that counts in this particular "challenge". To change that, go here, and click on "Edit PrimeGrid preferences" near the bottom. Then check the checkbox next to "321 Prime Search LLR (321)", uncheck the one next to "Generalized Cullen/Woodall Sieve (GCW-Sieve)", and save the changes at the bottom of the page.

If you do this you should watch your CPU temperatures. LLR is very similar to Prime95, so it runs your system about as hard as anything useful can. If you've overclocked that 6700k you may need to lower your overclock.

Good luck, and I hope you have fun.

Hey thanks for the heads up. I went ahead and updated my preferences like you suggested and will wait for BOINC to switch over to 321 Prime Search LLR (321). Right now it's still churning away on the Cullen/Woodall Sieve.
Thanks for the tip on watching my temp when BOINC switches over. I do have my 6700k oc'ed so I will keep an eye on the temps when it does.

 

[bbhaag]

If you do this you should watch your CPU temperatures. LLR is very similar to Prime95, so it runs your system about as hard as anything useful can. If you've overclocked that 6700k you may need to lower your overclock.

Good luck, and I hope you have fun.

Wow you weren't lying. Once it switched over the temps on my cpu skyrocketed. Core3 was hitting 100*C and the other three were not far behind. The only way I was able to get my temps under control while maintaining my oc was to take the side off my case. Once I did that temps settled in the low to mid 90's.

 

[StefanR5R]

@bbhaag, I followed the link to your host in @Ken g6's post and am seeing your first LLR-321 result reported with ≈2h25m run time. Sometimes the run times which are recorded in these tables are wrong (longer than the real run time), but I suppose this one is correct. Impressive, because this looks like you already mastered the secret science of performance optimization by means of app_config.xml text configuration.

Incidentally I have processors similar to yours; i7-7700K @ 4.4 GHz. I get run times of ≈2h15m from them with <cmdline>-t 6</cmdline>, and virtually the same with <cmdline>-t 7</cmdline>.

Comparing run times of different tasks is flawed though, because the workload of each task is a bit different. But before this PrimeGrid challenge started, I took some measurements on a Haswell Xeon E3 @ 3.4 GHz, always using the very same task for all configs, which makes these results directly comparable to each other with pretty high precision:

cmdline ... run time ... points per day
-t 8 ... 208 minutes ... 33,800 PPD (100 %)
-t 7 ... 209 minutes ... 33,600 PPD (99.5 %)
-t 6 ... 213 minutes ... 32,900 PPD (97.5 %)​

Of course, different host hardware means that run times, and thereby PPD, will scale somewhat differently with increasing -t parameter.

On another note, I see that your host still has some GCW-Sieve tasks queued up. If you want to concentrate on LLR-321 for the ongoing PrimeGrid challenge, you can abort the GCW tasks which you haven't already started (or even those which you started, but you won't get credit for the partial computation that you already did on these tasks). In contrast, you should not suspend the GCW tasks, because as long as you have suspended tasks in your queue, the client will not download new tasks when it runs idle; not even 321-LLR tasks.

 

[StefanR5R]

It's now taking more than just @StefanR5R to beat EVGA.

Soon it may take more than StefanR5R to beat the french cloud, even though it merely runs random subprojects at default = worst performance settings.

(from https://www.seti-germany.de/pg_challenge/stats/team/201802/132 -> "Vergleich")

 

[Ken g6]

it merely runs random subprojects

I think they've focused on 321 LLR. When @bbhaag started up, he(?) was doing Riesel Sieve by default.

 

[Ken g6]

Day 2 stats:

Rank___Credits____Username
4______2477261____xii5ku
25_____553108_____crashtech
50_____225482_____Howdy2u2
88_____127266_____Ken_g6
140____63646______zzuupp
153____53918___10esseeTony
161____48877______Orange Kid
215____24559______bbhaag
253____14675______MarcelliusIkaP
258____14651______SlangNRox
259____9815_______Kiska

Rank__Credits____Team
2_____7968072____Czech National Team
3_____7062888____SETI.Germany
4_____5573197____Sicituradastra.
5_____3613263____TeAm AnandTech
6_____3356555____L'Alliance Francophone
7_____2630656____Crunching@EVGA
8_____1846390____BOINC@Poland

In good news, welcome, @bbhaag! In bad news, it looks like the French are sure to pass us. We should also keep an eye on China. In weird news I did more work in a few hours yesterday on Xansons for COD than I did in two days for PrimeGrid.

 

[StefanR5R]

I think they've focused on 321 LLR.

The majority of the in-progress tasks on the french cloud are 321 LLR, but some of the in-progress tasks of other subprojects have been downloaded yesterday and today; some even just minutes ago. Therefore I suspect that the fact that they have more 321 tasks than of the other subprojects, is caused simply by the project server offering more 321s now, not due to operator intervention.

it looks like the French are sure to pass us.

Though I very much suspect that they do so unintentionally; involuntary even. ;-)

We should also keep an eye on China.

From the looks of their output curve, either these guys started a bit late (after new years celebrations?), or they run a lot of single-threaded tasks too.

Incidentally I have processors similar to yours; i7-7700K @ 4.4 GHz.

After monitoring these ones some more, I configured an AVX offset in the BIOS to get them down to 4.2 GHz. Even at 4.0 GHz their performance doesn't differ much from 4.4 GHz, where thermal throttling happened at times. That made me wonder how near I was to the point of getting invalid results.

That thick layer of thermal insulating paste that they have between die and lid is not a good match for workloads like PrimeGrid. Well, I purchased these primarily to feed my GPUs, not to run PrimeGrid LLR on them... (These i7s are actually sitting in waterloops together with 2 GPUs and 360 + 240 radiators. The GPUs are off now, the water is cold, yet the CPU dies still get hot. Good job, Intel.)

 

[TennesseeTony]

My output may not be exceptional (quad core Xeon v5), but at least I'm ahead of Michael Goetz by 5 places......

 

[ao_ika_red]

That thick layer of thermal insulating paste that they have between die and lid is not a good match for workloads like PrimeGrid. Well, I purchased these primarily to feed my GPUs, not to run PrimeGrid LLR on them... (These i7s are actually sitting in waterloops together with 2 GPUs and 360 + 240 radiators. The GPUs are off now, the water is cold, yet the CPU dies still get hot. Good job, Intel.)

Fortunately der8auer's office is still within your striking distance so I assume purchasing his delid tool won't cost much in shipping.

My output may not be exceptional (quad core Xeon v5), but at least I'm ahead of Michael Goetz by 5 places......

Where's your shiny new TR, Tony? Is it still doing WCG?

 

[Ken g6]

Day 3 stats:

Rank___Credits____Username
4______3776709____xii5ku
23_____872279_____crashtech
63_____299338_____Howdy2u2
99_____191229_____Ken_g6
127____127500_____zzuupp
171____78549___10esseeTony
175____78281______Orange Kid
178____73732______bbhaag
221____49083______Kiska
283____19598______MarcelliusIkaP
289____19547______SlangNRox

Rank__Credits____Team
3_____11316773___SETI.Germany
4_____9177004____Sicituradastra.
5_____6915061____L'Alliance Francophone
6_____5585850____TeAm AnandTech
7_____4061881____Crunching@EVGA
8_____3119050____Team China
9_____2918024____BOINC@Poland

We're still well ahead of EVGA, but China has polished off the Polish.

P.S. Thanks to Stefan for finding those charts.

 

[StefanR5R]

I just crunched a few ego-boosting numbers. Points are taken at March 19, 18:00 UTC.

xii5ku (TeAm AnandTech):

points = 3,761,915
CPU threads * GHz = 1033
points / (CPU threads * GHz * days) = 1214​

dthonon (L'Alliance Francophone):

points = 6,757,910
low estimate of the resources:

CPU threads * GHz = (1996 * 1.8 + 15 * 3.0) = 3638
points / (CPU threads * GHz * days) = 619​

high estimate of the resources:

CPU threads * GHz = (1996 * 2.6 + 15 * 3.0) = 5235
points / (CPU threads * GHz * days) = 430​

1.8 GHz is the AVX base frequency of E5-2699v4, 2.6 is its all-core AVX turbo. dthonon uses VMs on 1996 E5-2696Cv4 threads, and 15 threads on smaller machines, if I counted correctly. I don't know at which clocks they run. The fact that these are VMs would be beneficial, if other VMs on the same metal carry a lighter load than PrimeGrid LLR. dthonon runs singlethreaded 321-LLR and also some tasks of random other subprojects.

xii5ku uses about 90 % Xeon E5v4 and 10 % desktop processors in this challenge, and runs exclusively 321-LLR in multithreaded mode on all threads of each CPU ( - edit - on Xeon E5: several multithreaded tasks per processor).

Number of hardware threads times core clock is a bad metric for hardware resources, but the majority of the machines compared here are very similar to each other.

 

[TennesseeTony]

Where's your shiny new TR, Tony? Is it still doing WCG?

The AMD version of the AVX doesn't seem to be kicking in or something, while running 30 threads it was taking more than 5 hours per task, whereas the Xeon v5 I mentioned is completing a task in less than 4 hours with just 4 cores.
I finally completed all my Winter goals for WCG (as listed in the milestone's thread), so, I put it back to.....what am I running?....CSG I think.

 

[Markfw]

The AMD version of the AVX doesn't seem to be kicking in or something, while running 30 threads it was taking more than 5 hours per task, whereas the Xeon v5 I mentioned is completing a task in less than 4 hours with just 4 cores.
I finally completed all my Winter goals for WCG (as listed in the milestone's thread), so, I put it back to.....what am I running?....CSG I think.

Is it running Windows or Linux ? I will be doing both for TR number 4 (like all the rest)

 

[StefanR5R]

From what I read, LLR v3.8.20 (which PrimeGrid uses) uses the same instruction sets on Zen which it would use on the original Phenom and Phenom 2 processors (AMD family 10h), whereas Jean Penné's v3.8.21 uses AVX instructions on Zen but gets only at the order of 10 % better performance.

 

[ao_ika_red]

The AMD version of the AVX doesn't seem to be kicking in or something

whereas Jean Penné's v3.8.21 uses AVX instructions on Zen but gets only at the order of 10 % better performance.

Zen's AVX implementation is more or less like Bulldozer which still uses 2x128-bit FMAC unit instead of Intel's single 256-bit AVX unit so it won't be as good as Intel. That's probably why AMD CPU is never mentioned on PG's YoTD challenge page.
But based on my experience it's still better with crippled AVX than using SSE on supported apps.

Anyway, where's Jean Penne LLR custom app that mentioned by Stefan? I can't find it.
Edit: Found it. It was posted by Stefan but somehow I missed it.

 

[ao_ika_red]

Comparing run times of different tasks is flawed though, because the workload of each task is a bit different. But before this PrimeGrid challenge started, I took some measurements on a Haswell Xeon E3 @ 3.4 GHz, always using the very same task for all configs, which makes these results directly comparable to each other with pretty high precision:
cmdline ... run time ... points per day
-t 8 ... 208 minutes ... 33,800 PPD (100 %)
-t 7 ... 209 minutes ... 33,600 PPD (99.5 %)
-t 6 ... 213 minutes ... 32,900 PPD (97.5 %)Of course, different host hardware means that run times, and thereby PPD, will scale somewhat differently with increasing -t parameter.

The result is in. was following Stefan's experiment since yesterday and apparently, the result isn't good for Bulldozer-derived CPU. Due to its odd arch, losing one thread means additional 2 hrs duration per task. I had ~16 hrs on 4 thread two days ago, but now it's down to ~18 hrs.

 

[Ken g6]

Day 4 stats:

Rank___Credits____Username
4______5060459____xii5ku
21_____1232209____crashtech
74_____348779_____Howdy2u2
101____250448_____Ken_g6
139____166970_____zzuupp
165____123081_____bbhaag
175____113094___10esseeTony
182____107804_____Orange Kid
191____98154______Kiska
270____39222______SlangNRox
302____24492______MarcelliusIkaP

Rank__Credits____Team
3_____15716856___SETI.Germany
4_____12712466___Sicituradastra.
5_____10968413___L'Alliance Francophone
6_____7564717____TeAm AnandTech
7_____5467660____Crunching@EVGA
8_____4652899____Team China
9_____4071879____BOINC@Poland

Our 6th is looking solid. But @ao_ika_red is out of the individual points now unless he comes up with more CPU power.

 

[Howdy]

Well I'm dropping like a rock. Backed off everything on the offending machine and pulled RAM left 2 sticks that passed mem test........still throwing invalid results........it is staying offline until further review OR I beat it into a pile of scrap.

 

[ao_ika_red]

But @ao_ika_red is out of the individual points now unless he comes up with more CPU power.

Dang, 2 out of 7 finished tasks come as invalid. I've been throwing my kitchen sink since the day one and had nothing left. These challenges really need solid desktop CPU to perform, not just repurposed mobile CPU.

 

[Ken g6]

Dang, 2 out of 7 finished tasks come as invalid. I've been throwing my kitchen sink since the day one and had nothing left. These challenges really need solid desktop CPU to perform, not just repurposed mobile CPU.

My mobile CPU does fine as long as I keep its clock speed down with ThrottleStop. So maybe you actually need less CPU power.

 

[ao_ika_red]

These challenges really need solid desktop CPU to perform, not just repurposed mobile CPU.

My mobile CPU does fine as long as I keep its clock speed down with ThrottleStop. So maybe you actually need less CPU power.

What I mean is the Athlon 845 as repurposed mobile cpu (harvested Carrizo mobile APU). It doesn't have ample L2 and L3 cache like its FX older brother. Apology for the confusion.
My C2D laptop is still doing fine, though. It only needs new TIM after this week's challenge.

 

[StefanR5R]

Look at @bbhaag's output graphs — super regular, like clockwork!

 

[bbhaag]

Look at @bbhaag's output graphs — super regular, like clockwork!

Is that ok? I honestly haven't had a chance to check on it since I set it up the other day. I work very long hours(6am-6pm) and when I get home in the evenings I have to help with homework and dinner. By the time I get a chance to relax in evenings I'm asleep by 9pm.haha
I hear it chugging along over in the corner so I just assumed it was working ok.

 

[StefanR5R]

It's perfect. The periods between task completions are at the expected range for your hardware (with the proper app_config which you have since early on), and there are no spikes below zero in the hourly graph (which would occur if results came out invalid due to an error on your host, discovered by the project by check on two other hosts). Thumbs up for your reliable setup.