PrimeGrid Challenges 2021

[Ken g6]

Day 1 stats:

Rank___Credits____Username
11_____845115_____crashtech
23_____471209_____Icecold
34_____325243_____emoga
39_____301966_____biodoc
47_____242391_____xii5ku
75_____153819_____The Great Cornholio
82_____138500_____Orange Kid
91_____115262_____Ken_g6
143____70270______SlangNRox
146____63553___10esseeTony
159____56316______waffleironhead
160____56237______markfw
295____12557______Lane42
413____145________geecee



Rank__Credits____Team
4_____4434665____Czech National Team
5_____4111872____SETI.Germany
6_____3662986____AMD Users
7_____2852565____TeAm AnandTech
8_____2379118____Storm
9_____1338638____BOINC@MIXI
10____1284694____Sicituradastra.



I can't believe @Markfw is doing PrimeGrid!

Anyway, our rankings are much better when everybody's not doing some other project.

 

[Markfw]

I am doing 12 of these at a time in 5 minutes 10 seconds (approx). How good is that ?

Edit: surgery went ok, but it only has a 50% chance of working, or I need a bigger surgery. Off to bed to try and recover, very tired and in pain.

 

[StefanR5R]

For each number which is tested, two workunits are emitted: One which takes a few hours to compute (on modern CPUs), and one which takes a few minutes to compute.

Sometimes your computer receives long tasks, sometimes it receives short ones.

Not long ago, PrimeGrid had only the long type of WUs, and it went with a quorum of 2 like we know it from many other projects:

• For each number to test whether it is prime, the server generated one workunit.
• The server generated 2 tasks out of this WU, and assigned these tasks to two different hosts.
• Each of these two computers worked on its task, and when finished, it sent the server the result. The result could either be that this number is prime, or more frequently, that this number is not prime and that the calculation left a particular "residual".
• The server waited for both results to come in (or if one computer errored out, it would generate another task of this WU, assign it to another host, wait for this to return a result). As soon as it got two results, it would compare whether these were the same. (If not, it would generate another task, and so on, until there are two matching results.)

So in short, two computers needed to perform the very same hard work and needed to arrive at the very same result.

But during last year, PrimeGrid introduced an application update which reduced the required work:

• For each number to test whether it is prime, the server generates two workunits, one after the other. The first one is long, the second one very short and builds upon the first one.
• Of the first WU, just a single task is generated and sent to one host.
• This host performs the hours long classic work of the primality test. But it does not only return a result (it's prime, or it's not plus residual), it also returns several files in which state from several intermediate stages of the calculation is documented.
• The server takes the files and builds a new workunit from it. The purpose of the workunit will be to check if all of the documented calculation stages were performed correctly.
• A single task is generated from this WU and assigned to another host.
• This other host only needs to run a few very quick checks over the documentation data, and returns to the server whether or not it agrees.
• If yes, the result of the first, longer WU is deemed valid. If not, the server needs to generate another task out of the first WU and lets a third host have another go for a hopefully correct calculation.

That is, previously there were two hosts of donors calculating for several hours, typically in parallel. Now, there is one host calculating for several hours, and then another one spending just a few minutes for a check, in series.

Edit: This new method is implemented in many but not all of the various subprojects of PrimeGrid. The subprojects which have this new fast proof mechanism are the ones which are marked with an 'F' on the PrimeGrid front page, https://www.primegrid.com/. PPS-DIV is one of these modernized subprojects.

 

[StefanR5R]

I currently have just one 2xE5v4 in the race.

Yesterday when I came home from work, I surprisingly found another 2xE5v4 to join in on the Fermat Divisor Search.

(I now have 1.1 kW at PrimeGrid and 1.1 kW at SiDock, and it's a little more than I really need to heat the entire home. I've got several more space heaters which are switched off, and likely will stay off. I no longer want to have windows open 24 h/d in winter.)

 

[biodoc]

View attachment 37799

EDIT: I switched the 3950X to 8 tasks x 4 threads per task.

8 X 4 was an improvement for the 3950X @105 watts

 

[Icecold]

I really should have ran some tests on my hardware to determine optimal PPD before the challenge(As Stefan and possibly others had suggested to). I've been running with SMT on, 2 threads per task on my 3900x's and 3950x's. The task time on the 3950x's is significantly worse than the 3900x's with these settings so based on @biodoc 's post just now I changed those to 8 processes at 4 threads each for any new tasks downloaded on those. I also added 16 Intel cores / 20 threads last night(across multiple machines is the reason for the thread count not being double the core count). I'm hoping to make some headway against the 4 people ahead of me that are all grouped pretty close together.

We're way ahead of Noobs on this competition too, so that's good to see

 

[StefanR5R]

I really should have ran some tests on my hardware to determine optimal PPD before the challenge

As long as somebody else in the team provides good test results for the same processors, you don't have to test yourself. :-)

My own test results are of limited utility to the team, because my hardware is not particularly common. But I am content with contributing to the development of the testing methods. Some time ago I thought about getting myself some more diverse hardware to experiment with. But I soon dropped this thought and went for several computers built with same parts, for much easier management.

 

[biodoc]

I should have done more tests on the 3900X and 3950X. I put together a Zen 2 test guide for next time. Could you guys take a look to see if I made any errors?

 

[StefanR5R]

When we get to subprojects at which the FFT data exceed 16 MByte, we also need to test with fewer concurrent tasks than there are CCXs. My prediction is though that some overuse of CPU caches comes at a smaller performance hit than cross-CCX communications.

I should rewrite my script to offer a time-limited mode: Kill the test after a chosen run time, gather the progress percentage of all workers, calculate the average. This will make testing a lot easier to plan. If all tests run for a fixed period, it will also be easier to collect peripheral measurements during the tests, e.g. power meter readings if the power meter isn't an expensive one with logging.

 

[crashtech]

Hey @Ken g6 , I suggest the use of this link in the OP, so it doesn't have to be edited for every contest:

PrimeGrid

www.primegrid.com

Right now it points to the last challenge of 2020.

 

[Ken g6]

Day 2 stats:

Rank___Credits____Username
11_____1795727____crashtech
20_____1192724____Icecold
32_____734715_____xii5ku
38_____663503_____emoga
42_____621017_____biodoc
71_____356412_____Orange Kid
79_____321808_____Ken_g6
82_____317464_____The Great Cornholio
90_____289996___10esseeTony
153____149288_____markfw
156____144977_____SlangNRox
186____101315_____waffleironhead
190____98935______Lane42
429____3266_______geecee
492____296________Howdy



Rank__Credits____Team
4_____9962750____Czech National Team
5_____9581340____SETI.Germany
6_____8041659____AMD Users
7_____6791450____TeAm AnandTech
8_____5009572____Storm
9_____2986383____BOINC@MIXI
10____2842199____Crunching@EVGA



I guess I passed @Skivelitis2? Watch your back, @Orange Kid!

Hey @Ken g6 , I suggest the use of this link in the OP, so it doesn't have to be edited for every contest:

Thanks. Done.

 

[Markfw]

OK, I am bored, and I am about ready to get my covid vaccination, so I added 12(a 3900x) cores to this.

And my surgery is over, no bad news YET on the biopsy.

 

[Orange Kid]

@Ken g6 said "I guess I passed @Skivelitis2? Watch your back, @Orange Kid!"

@TennesseeTony is going to run us over first.

 

[Skivelitis2]

I don't mind getting passed. Reminds me of when I was a wee lad and my Dad praised me for being the best player on my Little League team. I told him I wish I was the worst player...just think how good we'd be! Could use some more tire tracks on my back. Pass away!

 

[TennesseeTony]

....I told him I wish I was the worst player...just think how good we'd be! ....

Wow. Deep. I like it!

@TennesseeTony is going to run us over first.

2 of 3 boxes were idle this morning, no idea how long. All settings were correct, but after a few updates to PG they finally got tasks. I am running 3 R9-3950X's, HT enabled, 3 threads per task, and also running Wall-Sun-Earth-Moon-Floor tasks on the GPU/remaining CPU. Getting about 3 hours per task on 2 of them, and 3 hours 15min on the one box I have neutered to 105w PPT.

Looking at biodoc's chart a few posts up, 10 tasks by 3 threads does not seem very efficient. I guess I will cut the GPU app out and go 8tsk by 4t soon.

 

[Icecold]

I like your outlook on that @Skivelitis2 . I have a strong feeling I might be getting run over tomorrow so I'll keep that in mind and try to have the same attitude towards it when I do

 

[VirtualLarry]

Ok, I'm finally in, sorry for the delay. Running 2x 3600 CPUs, each with 3x 4thread tasks. (*Not really sure if that's most efficient or if 6x 2 would be.)

 

[StefanR5R]

If it behaves similar to biodoc's 3700X, then

• 6 single-threaded tasks probably give highest throughput,
• 6 dual-threaded tasks would be a regression from that,
• 3 4-threaded tasks would be a regression in throughput but a win in run time.

Edit: Oops. See correction by @biodoc.

 

[biodoc]

3 4-threaded tasks would be a regression in throughput but a win in run time.

Four 3-threaded tasks would be a better fit since the 3600 has 6 threads /CCX vs 8 threads per CCX in the 3700X. See this post.

 

[Icecold]

(I just passed Honza, hopefully will pass the others)

Edit - I spoke too soon, Honza is back ahead of me

 

[Markfw]

OK, I bought a 3900x from a user here, and its running primegrid as a test. so total 3 3900x'x with 36 tasks, running about 5 minutes each.

I forgot to disable WCG on one of the three, so the last 24 hours are pretty much ONE. But the next update will be all 3.

 

[StefanR5R]

I have a strong feeling I might be getting run over tomorrow

You remained safe on Saturday, and in fact indulged in some overrunning yourself. However, tables may turn on Sunday. :->

 

[crashtech]

As if we don't know by now how much Stefan loves last minute surprises!

 

[Icecold]

I had the day wrong, but now I'm thinking tomorrow will be the day I get run over I might be able to run over 1 or 2 people from other teams tomorrow before it happens, though, hopefully.

 

[Ken g6]

Day 3 stats:

Rank___Credits____Username
13_____2770428____crashtech
17_____2158686____Icecold
21_____1869361____xii5ku
44_____996620_____emoga
45_____994509_____biodoc
55_____779680___10esseeTony
73_____587509_____Orange Kid
75_____562668_____Ken_g6
79_____510320_____The Great Cornholio
128____289005_____Lane42
140____267075_____markfw
163____213755_____SlangNRox
233____109054_____waffleironhead
255____91061______Howdy
312____53428______VirtualLarry
326____46873______Skillz
483____3266_______geecee

Rank__Credits____Team
4_____16792278___SETI.Germany
5_____15748277___Ultimate Chaos
6_____12729007___AMD Users
7_____12303307___TeAm AnandTech
8_____8089393____Storm
9_____4676270____Crunching@EVGA
10____4670175____BOINC@MIXI

That's quite a turnout! Looks like seventeen is not a bust.