The Scottish Boinc Team

so the next PG challenge is SR5 LLR

With headless workstations all over the house, turning off hyper-threading is not going to happen. Is it best to multi-thread or to run 1 thread per WU? Gut feeling is mt because they'll waste less time at end of challenge

I don't think hyper-threading hurts. The LLR apps run hot though so be sure you can keep your cool.

Thanks

I am running a few tests on 3900X at the moment, 4-thread WU running either 3 at a time, or 6 at a time. HT on in both tests so not a true test I guess

90-100mins for 4-thread WU, HT on but only 3WU running

Looking to be around 135mins for 6 x 4-thread WU.

I don't know if these are good scores or not - will try Xeon processors tomorrow

I think this means that Windows is faster http://www.primegrid.com/workunit.php?wuid=641917782

My wingman also seemed to have HT on but I can't tell if he was running 6 or fewer WU

The best way to run LLR tasks on a hyper-threaded machine is by using only half the threads (!!). It's just a smidgen slower than turning off HT ( so the experts say ). I run with HT on and just use half the threads. If you also run a GPU on the machine, its best if you run one that uses the CPU lightly, like collatz. Otherwise its Hobson's choice of giving a full core to the GPU or letting LLR and the GPU fight it out.

SR5 units are 2 million digits long so high muti-threads work well. LLR really likes L3 cache and suffers when it has to use regular old memory. With high digit counts, running multiple units can exhaust L3 and cause this condition.

I'll see if I can't find a thread wich discusses this better ..

http://www.primegrid.com/forum_thread.php?id=8901 briefly discusses this and has relatively recent table of FFT sizes
http://www.primegrid.com/forum_thread.php?id=9023 a discussion today on AMD vs Intel for LLR .. mackerel is always interesting to read.

and a personal note, dual cpu windows machines have troubles for me. I'll be trying Process Lasso this time to see if I can keep the units on one CPU or the other. Given all units have the same name, not sure it will work.

Gibson Praise wrote: ↑Sat Feb 08, 2020 1:02 pm The best way to run LLR tasks on a hyper-threaded machine is by using only half the threads (!!). It's just a smidgen slower than turning off HT ( so the experts say )

Thank you - that saves me a day's work going around with a monitor and keyboard. I also noticed that a 4-thread WU uses 3Gb ram.

I don't suppose there is any real way to tell how much L3 cache is being used - trial and error, eh?

Gibson Praise wrote: ↑Sat Feb 08, 2020 1:02 pm With high digit counts, running multiple units can exhaust L3

From what I am seeing though, running 8 or 16 threads per WU, neither halves the run time nor doubles the credits. Highest ppd I've seen so far with mt4, mt8, mt16 is for mt4. I think I'll try mt2 and see if my L3 cache can take it

Bear in mind that I am just a Points Whore - doesn't matter to me if I find the Prime or not (although I do reckon wingman should get equal recognition for any primes found)

davidBAM wrote: ↑Sun Feb 09, 2020 1:00 pm Bear in mind that I am just a Points Whore - doesn't matter to me if I find the Prime or not (although I do reckon wingman should get equal recognition for any primes found)

Go for it. I'd like to know how it turns out, definitely. PG has a strong cultural bias toward finding the prime!. The advice on the forums most always takes that as a given.

Tangent #1: February is the Tour de Prime at PG where the whole idea is to find t5k (top 5000) primes and for the first time in years I found one. Nifty badge and all! ( and yes, running GFN-16 gpu to get it was horribly inefficient and point-poor )

Tangent #2: I wonder what the Venn diagram of Post Whores and Points Whores looks like?

Are you trying to tell me that I post too much?

davidBAM wrote: ↑Sun Feb 09, 2020 6:34 pm Are you trying to tell me that I post too much?

Never!

On my dual CPU 14c/28t Xeon V4 rigs, linux Mint 19.1 is slightly faster than Win7 64 and the sweet spot for these chips appears to be 7 threads per WU. I started at 28t per WU, then 14t, 7t and 4t. Pay per thread increased to 7t per WU, then the pay per thread dropped running 4t per WU.

How many 7t threads were you running simultaneously please? And how does the PPD look?

8x 7t WU at a time but I appear to have had an error in the formula. Re-testing.

Okay, so HT on and hitting L3 cache as hard as you can.

I am scrapping all my notes and tackling it from the other end. HT still on but only very low counts of 4t WU until I get an idea of what each machine is capable of time-wise

Note to self: WU downloaded on 11-Feb fit this description

Dual Xeon V4 14c/28t 3Ghz, HT on

Thank you. What is approx run time on "linux 7 8" please

2990WX : Average at 4 x 4t: 12067 secs (2139pts / WU)

3900X : Average at 3 x 4t : 6251 secs ( 2249pts / WU)

asus2: Average at 2 x 4t: 8415 secs
t7910: Average at 2 x 4t: 9715 secs

asus1: 9614

davidBAM wrote: ↑Tue Feb 11, 2020 2:32 pm Thank you. What is approx run time on "linux 7 8" please

Runtimes range from 2.1 to 2.9 hrs with avg runtime of 2.62 hrs

Hmmm - I think my 2990wx machines might be throttling back due to heat. What is their max running temperature before that happens please?

I think it is 65C. I have to down clock mine to stock (3.0) to run PG. @3.3 they throttle ... on air.