PrimeGrid Challenges 2021

[Ken g6]

Preliminary final stats:

Rank___Credits____Username
43_____18132000___crashtech
48_____16932000___xii5ku
51_____15396000___biodoc
56_____12852000___emoga
57_____12828000___Orange Kid
82_____7908000____The Great Cornholio
103____5352000____Lane42
128____3828000____SlangNRox
227____1332000____Ken_g6
274____672000_____waffleironhead
393____72000______geecee

Rank__Credits____Team
6_____117648000___Microsoft
7_____116664000___Storm
8_____99504000___Save The World Real Estates
9_____95304000___TeAm AnandTech
10____80772000___AMD Users
11____79788000___UK BOINC Team
12____71676000___L'Alliance Francophone

Well, still top 10, anyway.

 

[StefanR5R]

Thanks for the stats.
This one was an uphill battle for those without Turing and Ampere cards.

 

[Orange Kid]

Thanks for all the stats.

 

[Skivelitis2]

Thanks for the stats.

 

[Skillz]

I shall start participating here since I'm not longer participating in Formula-BOINC.

I think we are overall 6th. Let's see if we can't improve that.

 

[crashtech]

We have a few weeks to bone up on our ESP-LLR skills.

 

[StefanR5R]

Now we have one week to bone up on our ESP-LLR skills. :-)

From PG's front page:
llrESP is now at 5.5 M digits, and a prime would appear in 17th place of T5K. This is just after llrCUL and llrWOO, and ahead of llr321.

From the subproject primer:
However, making it big in T5K is not the real goal behind the llrESP project. Rather, the llrESP project looks for prime numbers of the form k*2^n+1 for nine particular ks (but potentially unlimited ns), with this goal in mind: As soon as a prime for one of these ks is found, the conjecture that this k is not a Sierpiński number would be proven. In this event, the search would abandon this k and carry on with the remaining few of this set of ks.

From PG preferences:
The recent average CPU time is ~58 hours. A 10 % "long job" credit bonus is given to results in this subproject. And for the above mentioned reason, a 10 % "conjecture" credit bonus is put on top.

From apps.php:
llrESP is at version 9.01 (mt), installed on Dec 22, 2020.

From client_state.xml of a Linux x86-64 computer which downloaded llrESP today:
The current LLR v2 binary is sllr2_1.1.0_linux64_201114, and its download URL is http://www.primegrid.com/download/sllr2_1.1.0_linux64_201114.gz.
If you don't have it in a boinc-client already, you can download it for testing:
wget http://www.primegrid.com/download/sllr2_1.1.0_linux64_201114.gz
gunzip sllr2_1.1.0_linux64_201114.gz
chmod +x sllr2_1.1.0_linux64_201114

The ranges of k and n of llrESP tasks which are currently in progress is listed at PG's subproject status page. I examined these k and n on a Haswell CPU and on a Zen 2 CPU, and got the following FFT lengths on both CPUs. Note, the LLR2 program may choose different FFTs on CPU architectures with different vector command set support.

k in progress	n in progress	FMA3 FFT length	FFT data size
91,549​	17,723,310...18,318,630​	1728K...1728K​	13.5...13.5 MB​
131,179​	17,891,498...18,319,142​	1728K...1728K​	13.5...13.5 MB​
163,187​	17,907,279...18,318,903​	1728K...1920K​	13.5...15.0 MB​
200,749​	17,950,242...18,318,786​	1920K...1920K​	15.0...15.0 MB​
202,705​	17,838,820...18,319,094​	1920K...1920K​	15.0...15.0 MB​
209,611​	18,060,920...18,318,584​	1920K...1920K​	15.0...15.0 MB​
227,723​	17,929,429...18,318,661​	1920K...1920K​	15.0...15.0 MB​
229,673​	17,758,149...18,319,137​	1920K...1920K​	15.0...15.0 MB​
238,411​	17,888,784...18,319,020​	1920K...1920K​	15.0...15.0 MB​

 

[Skillz]

We got approximately 5 minutes before this starts.

Make sure you select ONLY Extended Sierpinski Problem LLR (ESP) to run on your CPU and don't start downloading the tasks until roughly 5 minutes from this post. Their is a timer on the PG web site to let you know exactly when to start.

PrimeGrid

 

[Skillz]

23 hours and 15 minutes into the challenge.

Rank	Name	Score	Tasks
1​	Ural Federal University	3 579 823.52​	233​
2​	TeAm AnandTech	3 487 130.93​	282​
3​	Antarctic Crunchers	3 255 881.38​	255​
4​	Czech National Team	2 914 177.83​	208​
5​	SETI.Germany	2 178 337.82​	200​

We are looking pretty good so far. Anyone else wanna jump in to help close the gap on Ural Federal University?

2	Skillz	TeAm AnandTech	1 727 677.30	113
6	crashtech	TeAm AnandTech	798 120.42	75
15	biodoc	TeAm AnandTech	379 953.72	31
23	emoga	TeAm AnandTech	251 298.09	17
30	Orange Kid	TeAm AnandTech	181 668.14	10
55	Skivelitis2	TeAm AnandTech	74 135.11	21
83	Ken_g6	TeAm AnandTech	54 530.27	3
141	SlangNRox	TeAm AnandTech	18 196.27	1
171	Lane42	TeAm AnandTech	1 266.41	9
181	xii5ku	TeAm AnandTech	285.21	2

 

[Skillz]

Can anyone explain this one?

One guy, has two computers; 2600k and a Q9400. Combined they are doing over 3.8M points per DAY.

Computers belonging to Pavel Atnashev

How?

 

[emoga]

It has been addressed on the forums multiple times. Here is a post from Michael Goetz explaining it.

He’s got a full rack of servers in his apartment, with custom BOINC software and a custom version of LLR to get them to run without disk drives. They talk to the PrimeGrid server through some small gateway computers. Those gateways are what you see in his host list. They are not the computers doing the actual crunching.

He is also the person who wrote our LLR2 software.

 

[Skillz]

That's what I figured was happening. I did go looking around on the forums to see how he was doing it, but couldn't find anything. I did, however, discover that he was the guy who wrote the LLR2 software.

Guess I can cancel the eBay orders for a rack of Q9400s I was about to buy.

 

[StefanR5R]

23 hours and 15 minutes into the challenge.

Which is not much. — Back when they planned the 2021 challenge season, 3-vcore VMs at Digital Ocean were still in fashion. IIRC pricing has changed over there, and there might be a few guys out there who are now hooked on 1-vcore VMs. With ESP-LLR, and 5 days contest duration, that should work out to 1 result per VM in the contest (not counting those small verification tasks).

Here is a post from Michael Goetz explaining it.

I haven't looked for other posts on this topic, but to me, this setup doesn't look like it's because the cluster nodes are diskless. (You can put regular boinc and regular sllr2 on a networked filesystem.) Rather it looks to me as if this is about isolation of the cluster nodes from the internet.

 

[Endgame124]

Which is not much. — Back when they planned the 2021 challenge season, 3-vcore VMs at Digital Ocean were still in fashion. IIRC pricing has changed over there, and there might be a few guys out there who are now hooked on 1-vcore VMs. With ESP-LLR, and 5 days contest duration, that should work out to 1 result per VM in the contest (not counting those small verification tasks).



I haven't looked for other posts on this topic, but to me, this setup doesn't look like it's because the cluster nodes are diskless. (You can put regular boinc and regular sllr2 on a networked filesystem.) Rather it looks to me as if this is about isolation of the cluster nodes from the internet.

And if you’re aggregating the work though a single system it’s more entertaining if you use low power gear. I would totally use a raspberry pi or a athlon 64 myself

 

[crashtech]

From the specs, it's whatever they had lying around.

 

[Ken g6]

Day 2 stats:

Rank___Credits____Username
3______5412993____Skillz
8______2102364____crashtech
10_____1779906____xii5ku
23_____901874_____biodoc
24_____896008_____emoga
41_____441371_____Orange Kid
59_____276365_____Icecold
74_____219688_____Skivelitis2
89_____168584_____Lane42
101____147066_____Ken_g6
199____36513______SlangNRox

Rank__Credits____Team
1_____12382738___TeAm AnandTech
2_____12102730___Antarctic Crunchers
3_____10185969___Ural Federal University
4_____9764931____SETI.Germany

Wow, we're in 1st!

Which is not much. — Back when they planned the 2021 challenge season, 3-vcore VMs at Digital Ocean were still in fashion. IIRC pricing has changed over there, and there might be a few guys out there who are now hooked on 1-vcore VMs. With ESP-LLR, and 5 days contest duration, that should work out to 1 result per VM in the contest (not counting those small verification tasks).

So it may not stay that way without a lot more work.

 

[Skillz]

I'm giving her all she's got, Captain. -Scotty

 

[crashtech]

One of my 3700X rigs is spontaneously rebooting, so it's out of the race until I can fix it.
It's hot here.

 

[lane42]

Back in 1st

One of my 3700X rigs is spontaneously rebooting, Power supply ?

 

[crashtech]

I don't think so, that rig has been having problems for a while, and I already swapped out the PSU. It passes Memtest overnight but then gets errors when testing Primegrid projects. The logfiles from the Primegrid tests are actually a pretty handy way to tell if the system is stable. I thought I had it fixed by turning down the memory speed because the testing errors went away, but no so.

 

[Ken g6]

Day 3 stats:

Rank___Credits____Username
3______10392309___Skillz
9______4109063____xii5ku
10_____3489019____crashtech
25_____1464107____biodoc
28_____1307173____emoga
44_____667755_____Orange Kid
48_____630248_____Icecold
74_____369919_____Lane42
79_____349946_____Skivelitis2
98_____258600_____Ken_g6
239____36513______SlangNRox

Rank__Credits____Team
1_____23074656___TeAm AnandTech
2_____20583830___Antarctic Crunchers
3_____18947857___SETI.Germany
4_____16585266___Ural Federal University

@Skillz must be on cloud 9 right now! (Out of how many I'm not sure. )

 

[Skillz]

What? I couldn't cancel those Q9400s I bought on eBay so I put them together. They're rocking.

 

[Ken g6]

Day 4 stats:

Rank___Credits____Username
3______16971475___Skillz
8______7011617____xii5ku
11_____4772637____crashtech
25_____2013801____emoga
27_____1947124____biodoc
44_____1110631____Icecold
50_____892501_____Orange Kid
74_____553178_____Lane42
81_____499527_____Skivelitis2
104____353063_____Ken_g6
272____36513______SlangNRox

Rank__Credits____Team
1_____36162073___TeAm AnandTech
2_____31402692___Antarctic Crunchers
3_____27596099___SETI.Germany
4_____27094242___Team China

So, who's defying their power company to keep running today?

 

[StefanR5R]

I haven't watched at which point it happened, but quite a few workunits went up from 1920K to 2M FFT length = from 15 MBytes to 16 MBytes FFT data now (on computers on which LLR2 uses FMA3). 23 % of the tasks which I have running now are of the latter kind.

This won't matter on my Broadwell-EPs, they have got enough headroom. On my Zen 2 computers with their 16 MBytes cache segments, these 2M tasks apparently take only a little bit longer than the 1920K — but that's while I am going with a 1:2 ratio of concurrent tasks to core complexes. The throughput optimum with 1920K WUs was actually a 1:1 ratio if strictly cache-aware task scheduling is implemented. (But I went with 1:2 for shorter task durations.) Maybe I'll re-test with a 16 MB WU later.

 

[Ken g6]

More-or-less final stats:

Rank___Credits____Username
3______22528554___Skillz
8______9365689____xii5ku
14_____5994361____crashtech
28_____2478944____biodoc
31_____2199568____emoga
41_____1526308____Icecold
53_____1104035____Orange Kid
83_____612199_____Skivelitis2
84_____608260_____Lane42
107____445425_____Ken_g6
291____36513______SlangNRox

Rank__Credits____Team
1_____46899861___TeAm AnandTech
2_____39640205___Antarctic Crunchers
3_____39559505___SETI.Germany
4_____39214531___Team China

Wow, that's a decisive win!