So far I had run SGS and 321 sieve alongside. After reaching 321 sieve ruby, I deactivated it and focused solely on SGS.

Surprisingly, the CPU time for each SGS task went from approximately 970 s average to 1500 s average. I reactivated 321 sieve after 24 h and averages went down to the prior value. I think you can publicly view my tasks and see for yourselves, 321 sieve was off for May, 2nd. Recently I changed from 321 sieve to extended sierpinski and apparently SGS doesn't benefit from it. This is quite peculiar...

This increase happened for both machines (i7-4790K and i3-2120) though it was less pronounced for the i3, something like 1500 s vs 2000 s.

I guess this has to do with hyperthreading? Switching to fewer simultaneous tasks doesn't seem like a good solution, since that would probably leave CPU power unused. So should I keep running two subprojects?

Has anyone experienced this as well?


edit: I'm starting to believe this is more of a problem of scare ressources. Both machines have enough of ressource X to keep 2 or 4, respectively, SGS tasks running efficiently. 321 sieve doesn't need ressource X and if it runs alongside, no problem. But if I switch off 321 sieve the number of parallel SGS tasks increases, making the machine lack in ressource X (whatever that is) and CPU time per task increases.

Extended Sierpinki apparently requires ressource X as well, as running it parallel to SGS causes the increase as well.

LLR doesn't benefit from hyperthreading, while sieves do benefit from hyperthreading.

If you have hyperthreading enabled, this behavior makes sense.

If you want to run SGS, simply disable hyperthreading, or, equivalently, set BOINC to use only 50% of the "CPUs". Despite what you might think, you're *not* leaving processing power on the table when you do this. Hyperthreads aren't "full CPU cores", and LLR uses the part of the CPU that is NOT duplicated in the hyperthreads. There's only one full arithmetic unit per core (i.e., there's 1 arithmetic unit per 2 hyperthreads), so the best you can do with LLR is one task per physical core (i.e., 1 task for every 2 hyperthreads.)

If you leave hyperthreading enabled, you'll run twice as many SGS tasks simultaneously, but they'll take AT LEAST twice as long to run. They'll probably be slower, actually. The number of tasks you complete per day will either stay the same or go down -- but you will consume a lot more power and produce more heat.

The same or slightly less throughput, with more power consumption -- it's a lose-lose scenario.
____________
My lucky number is 75898⁵²⁴²⁸⁸+1

That still doesn't really explain while SGS took about 60% of the CPU time when 321 sieve ran simultaneously, or does it? Or is the important part that no more than 2 SGS (on a 2 physical cores system) tasks run at the same time?

What I meant with wasted CPU time:

Scenario 1: 4 SGS tasks with 1500 s each = 9.6 SGS tasks per hour

Scenario 2: 2 SGS tasks with 970 s each plus 2 321 sieves with 3600 s each = 7.4 SGS tasks per hour plus 2 "321sieves" per hour

Scenario 3 (no HT): 2 SGS tasks with 675 s each (just a guess, assuming it's 50% of HT enabled) = 10.6 SGS tasks per hour (but no 321 sieves)



My experience was that switching HT off did not halve the CPU time of SGS tasks. I didn't run it long enough to be sure, but I will do some testing and report back.

That still doesn't really explain while SGS took about 60% of the CPU time when 321 sieve ran simultaneously, or does it? Or is the important part that no more than 2 SGS (on a 2 physical cores system) tasks run at the same time?

Admittedly, not much testing done so far, but after about 30 SGS tasks on the i3, I have an average of 1170 s per task while with HT it is about 1900 s. So overall throughput would be better with HT on, at least on that machine and if this keeps on.

Tomorrow morning I'll see what happened overnight.

i3-2120 Win 10 x64:


50% processors usage (i.e. HT off) for 24 h (computer not used)

Throughput: 147 WUs
min CPU time: 1150 s


100% processors usage (i.e. HT on) for 24 h (computer not used)

Throughput: 162 WUs
max CPU time: 1963 s


Throughput has increased by 10%. I added min and max CPU times to show that even the slowest HT=on task took less than twice what the fastest HT=off task took.

From these results - and a short, couple of hours test I ran on an i7 some time ago - I would say, at least the current SGS tasks do benefit from hyperthreading.

Since it seems common knowledge here that SGS suffers from HT, are there statistics available?

From these results - and a short, couple of hours test I ran on an i7 some time ago - I would say, at least the current SGS tasks do benefit from hyperthreading.

Since it seems common knowledge here that SGS suffers from HT, are there statistics available?

I simply can't run all 8 threads on SGS with my old Dell PC's (4790) The temps hit 98 degrees and the fan screams!
Drop CPU to 50% and just use the 4 cores and temps are still in the high 80's/low 90's and no screaming fan. I'm comfortable with that.

I may try on a proper PC with better cooling..... or I might just stick to no H/T and not torture my system :)

1
I have a Dell 4770 (Optiplex 9020) and a Dell 4790 (T1700) both have similar temps. The 4770 I actually fitted a Noctua NH-L9i Low Profile heatsink which is much better than the crappy Intel one. It dropped temps a bit and is much quieter but we are still talking mid to high 80's.
These things only have a 80mm fan in the front...that's it. No provisions for more fans.

Good that you've already done the repaste, I like to be thorough in my posts about cleaning in case anyone non-technical wanders by, or is like my grandmother who put her PC in a closed cabinet because it was getting too loud.

Those temps are still too high, IMO, even in a 25C room. I have a couple crunchers in a closed room, and while they warm it up to 30C, they still won't get that hot.

If you aren't too worried about the longevity of the machines, it's not the end of the world as long as the fan noise doesn't bother you too much (and you can always replace them with better fans that move more air with less noise), but you probably won't see the full potential of the CPUs due to throttling. While the L9i is better than the stock cooler, it has a pretty low capacity. Have you taken the side panels off for more cool air access? I don't know how much ducting is going on inside those models, but it is an easy way for more direct cool air access.

As an extreme measure, especially since it's winter, you can always take off the panels, wrap them in bug netting, and put them outside under shelter...
____________
Eating more cheese on Thursdays.

That is one thing that that could be causing an issue that you mentioned Grebuloner.... ducting. Neither of these units has ducting directing air from the front 80mm fan to the CPU.
I know one of them DID have ducting but again, you can't fit a Noctua fan with all that plastic crap everywhere :)

I guess if you only have 1 fan, that ducting directing air over the CPU would be important :)

That could be the reason.

Oh well...as long as I don't have a fire! (Although I did buy an extinguisher the other day...just in case!) Dry powder of course.