PrimeGrid Challenges 2023

[Markfw]

OK, I have to go to the emergency room and not sure when I will be back, so I will lose some, but starting the fleet now.

 

[Orange Kid]

OMG, take care of yourself first. Computing can wait. Best to get well and rest. There is always the next one.
Take care, hope all will be well.

 

[Markfw]

Extreme sciatic pain

 

[Markfw]

OK< I learned something new. sciatic pain, or acute sciatic pain that I have . There is NO cure, just pain meds that are worthless. 8mg of morphine, and I am still in horrible pain.

So I am back, and aborted all the tasks. Will start them back up in one hour.

 

[Markfw]

OK, in 5 1/2 hours, I am in first overall, and Team Anandtech is first in the competition as a team. That didn't take long.

This is why I really wanted to be here for the start. Glad I made it.

 

[mmonnin03]

Damn, my 7950 got tasks too soon. So 6+ hours wasted.

I can't say my sciatic pain was extreme by any means but it sucks that the thing that helped it also hurt it in the short term. Hamstring stretching.

 

[Markfw]

Kind of ahead a little

Might be due to this guy

 

[Markfw]

But I have a problem. 2 of my 9554's are both on windows and the same version of lasso; First computer

Second computer:


It looks very different, and the scores and efficiency are WAY off. The second is WAY better. Why ??? The rules were created the same way !!!

 

[StefanR5R]

The green-on-black bar graphs of the second computer look like the first, hinting that all cores seem to be utilized evenly.

But the "all processes" table shows something unexpected:
Three tasks are sitting in the same 0…14 CPUs cluster, one in 16…30, one in 32…46, two in 48…62.

The affinities of the first computer are not entirely correct either:
One task is bound to all even-numbered CPUs, instead of just eight of those.

I don't know Process Lasso and its capabilities, therefore can't suggest the perfect rules. I will try to come up with an alternative for Epyc 9554 on Windows later today.

 

[StefanR5R]

Oh wait, the actions logs of both computers indicate to me that the rules in use are *not* what is desired (constraining tasks to CCXs). But as I said, I don't know Lasso, can't suggest better rules.

 

[StefanR5R]

[double post… my first one here]

 

[Markfw]

Oh wait, the actions logs of both computers indicate to me that the rules in use are *not* what is desired (constraining tasks to CCXs). But as I said, I don't know Lasso, can't suggest better rules.

it looks like it may have fixed itself. Will look again in the morning. Just my observing 9554-2 and 9554-3, both close to 8 hours now.

Thanks for looks Stefan ! I have to use this in windows.

 

[StefanR5R]

I think the rules happen to do something which is equivalent to what we want much of the time, but not always. Better would be to have rules which just do exactly what we want. :-)

 

[Markfw]

No stats for day 1 ?

 

[Ken g6]

No stats for day 1 ?

Too early for me. And I figured you did Day 0.5 stats. I'll aim for Day 1.5.

 

[StefanR5R]

@Markfw, one other thing which is wrong on the screenshot of the second computer is that the PrimeGrid tasks are listed at "above normal" scheduling priority. They should be "normal" at most, better just "idle" like in the screenshot of the first computer. BOINC default for CPU-only work is "idle" scheduling priority.

But the gravest problem is that Lasso forced more than one task on the same set of CPUs, whereas we want exclusive CPU sets for each task.

 

[Ken g6]

Day 1.5 stats:

Rank___Credits____Username
2______8492693____markfw
3______4674967____xii5ku
5______3415853____crashtech
9______2394788____w a h
21_____750663_____cellarnoise2
29_____535710_____Orange Kid
32_____495624_____biodoc
55_____278167_____Fardringle
56_____277126_____waffleironhead
107____129901_____Icecold
114____124927_____mmonnin
119____121890_____Letin Noxe
126____119108_____Ken_g6

Rank__Credits____Team
1_____21811424___TeAm AnandTech
2_____13813944___BOINC@AUSTRALIA
3_____10023048___Czech National Team
4_____7654644____Antarctic Crunchers

Nothing spooky about this challenge so far...except maybe for other teams.

 

[Ken g6]

Day 2.5 stats - half way already!

Rank___Credits____Username
2______15708330___markfw
3______9151505____xii5ku
5______6698072____crashtech
10_____4317701____w a h
15_____2497967____cellarnoise2
32_____948993_____Orange Kid
34_____933990_____biodoc
51_____630997_____Fardringle
57_____553623_____waffleironhead
68_____475458_____[TA]Skillz
108____303768_____Letin Noxe
119____249354_____mmonnin
121____249106_____Ken_g6
130____224004_____Icecold
293____501________DROFFUNGUS

Rank__Credits____Team
1_____42943377___TeAm AnandTech
2_____25925679___BOINC@AUSTRALIA
3_____22465546___Czech National Team
4_____15880842___Antarctic Crunchers

Welcome to DROFFUNGUS! 👨‍⚕️?

 

[cellarnoise]

The Fungus Amongst? Hopefully the best of us! And not the Last!

 

[crashtech]

Fungus humongous among us?

 

[Ken g6]

Day 3.5 stats:

Rank___Credits____Username
2______22885846___markfw
3______13388466___xii5ku
6______9866024____crashtech
11_____6188111____w a h
15_____3465976____cellarnoise2
33_____1414192____biodoc
36_____1367926____Orange Kid
44_____1211827____[TA]Skillz
61_____791603_____Fardringle
64_____777261_____waffleironhead
112____432029_____Letin Noxe
113____431746_____mmonnin
126____343303_____Ken_g6
130____319705_____Icecold
269____31083______DROFFUNGUS

Rank__Credits____Team
1_____62915107___TeAm AnandTech
2_____39232932___BOINC@AUSTRALIA
3_____32613346___Czech National Team
4_____25164306___Antarctic Crunchers

Australia doing well. I suppose it's the time of year when it's cool-ish in both hemispheres.

 

[Markfw]

Australia doing well. I suppose it's the time of year when it's cool-ish in both hemispheres.

think its more Nick using hundreds of computers from the cloud

 

[StefanR5R]

So this one is one of those challenges into which I invested a bunch of several (quick but quite definitive) optimization tests upfront. Of course it's the first time that I have Zen 4 in the game, so that one was worth testing a bit. I even tested Broadwell-EP once more, even though this one is comparably straightforward to configure both on the hardware platform level and the software level. Quite some testing was concerned with Zen 2, which is handicapped in llrESP by the small size of the CCXs.

After the fact, two items came to my mind, or were brought to my attention elsewhere, which I haven't explored yet:

• Especially on Zen 2, would it be worthwhile to configure logical CPU affinity not just on the level of a whole task, but for each single thread of a task?

• On Zen 4, what if AVX-512 is disabled in the BIOS? The way AMD implemented it, it theoretically comes with chances of good benefits over AVX2, but also with a certain potential of regressions. (link to a link)

The latter can be figured out almost as quickly as the computer can be rebooted (which isn't quite very quick), so I might check this out some time after the challenge. The former needs a bit more work.

 

[crashtech]

I always look forward to your insights, Stefan. I read that info that you linked, and I didn't understand all of it. Initially I came away with the impression that there were no significant downsides to AMD's implementation, though it appears at first glance to be a bit of a kludge. The part that I found most surprising was that afaiu running AVX-512 code doesn't incur a power/heat penalty on Zen4 the way it does on Intel.

 

[Markfw]

So this one is one of those challenges into which I invested a bunch of several (quick but quite definitive) optimization tests upfront. Of course it's the first time that I have Zen 4 in the game, so that one was worth testing a bit. I even tested Broadwell-EP once more, even though this one is comparably straightforward to configure both on the hardware platform level and the software level. Quite some testing was concerned with Zen 2, which is handicapped in llrESP by the small size of the CCXs.

After the fact, two items came to my mind, or were brought to my attention elsewhere, which I haven't explored yet:

• Especially on Zen 2, would it be worthwhile to configure logical CPU affinity not just on the level of a whole task, but for each single thread of a task?
• On Zen 4, what if AVX-512 is disabled in the BIOS? The way AMD implemented it, it theoretically comes with chances of good benefits over AVX2, but also with a certain potential of regressions. (link to a link)

The latter can be figured out almost as quickly as the computer can be rebooted (which isn't quite very quick), so I might check this out some time after the challenge. The former needs a bit more work.

And it appears that even with avx-512 off, it beat SR, with it on, it just blows it away.


I looked some more to be sure I was not cherry picking