Folding Forum

P16983 base 55,000. I don't think I've had such a big cpu project before.
My client is estimating a time of just under 12 hours.

Let those WU's hit that 30 core CPU slot on my 3950x system.... I'll appreciate those points.

Paul

It's currently running on 21 of my 24 (the rest are given over to the gpu and OS).
It's estimating my PPD at 285K, which is about typical for my cpu.
It just means it's one big project instead of lots of little projects for roughly the same points.

I just hope Voelz was able to set a lower limit on the cpu this monster is sent to.

aetch wrote:I just hope Voelz was able to set a lower limit on the cpu this monster is sent to.

More likely to be a higher limit set for the number of CPU threads. With 23,400 atoms in the WU, beyond a point using more threads does not increase the processing speed. If this project was using Core_A7 instead of Core_A8 that is being used, domain decomposition errors could also be an issue.

I don't see any entries for this project on my systems, but would guess these WUs have a large number of steps to get the number of points involved instead of being large in terms of atom count.

10 Million steps.

A quick scan of my logs shows previous work units up to 2.5 million steps.

Anyway, 4 hours in, no hiccups. It's on target for a sub 12 hour run. *fingers crossed* I haven't just jinxed it.

I've been consistently getting ones with over 200,000 base credit. Takes 1.3 days to complete with a 5900X.

More likely to be a higher limit set for the number of CPU threads. With 23,400 atoms in the WU, beyond a point using more threads does not increase the processing speed. If this project was using Core_A7 instead of Core_A8 that is being used, domain decomposition errors could also be an issue.

Is there any consensus regarding at what point/at what number of CPUs does the use of more threads not increase the processing speed? This would of course be useful information so that those excess cores are not "spinning their wheels" with lost efficiency.

Thanks!

rbpeake wrote: ↑Thu Feb 02, 2023 2:11 pm

More likely to be a higher limit set for the number of CPU threads. With 23,400 atoms in the WU, beyond a point using more threads does not increase the processing speed. If this project was using Core_A7 instead of Core_A8 that is being used, domain decomposition errors could also be an issue.

Is there any consensus regarding at what point/at what number of CPUs does the use of more threads not increase the processing speed? This would of course be useful information so that those excess cores are not "spinning their wheels" with lost efficiency.

Thanks!

In general there is no consensus as it depends largely on the size of the system being looked at by a particular project. There have been projects where more than 8 did not bring about any increase in speed, at least one project was tested on a monster system and still saw performance improvements at 128. Most recent and current projects see speedups with 16 or more CPU threads, but some definitely reach a "knee" in the speed versus thread count when it reaches 32. Other projects still see improvements past that, but there is limited availability of high thread count systems to test all projects past that.

They do try during internal testing to identify projects that see no improvement in processing speed past a certain number of CPU threads. So there have been projects released with a maximum of 16 or 24. At the same time some projects get a minimum thread count, most systems with that many threads available will be able to complete within the timeout.

So if you have a large system and notice a particular project doesn't run well at say 40 threads and takes about the same time at 32, report it and that info will be given to the researchers to help set assignment limits.