18401 - Optout

[aetch]

Simple question - how do I opt out of 18401 and other similar super massive CPU projects I see in the pipeline?

[Neil-B]

There isn't a project opt out (would cause issues with cherry picking) however if you report what issue the larger wus are causing it may be possible for adjustments to the assignment logic (say restricting to cpu slots of more than 4 cores) to be made ... if it is a download/upload size issue causing bandwidth/usage limit issues then they may be other options ... efforts can be made to persuade researchers to size wus better if the science allows for instance a really large atom count can be mitigated by fewer steps per wu keeping the overall folding time acceptable.

18401 is not a big wu re atom count but from the timeout, deadline and base credit I guess it may have a large number of steps which some science requires ... the longer deadlines and high bc are probably intended to allow for this.

If you can clarify what is causing you to wish to optout this project I'll feed it back to the researcher.

Historically there was a flag that helped manage certain types of big - I am not sure if it is still active/effective or whether it would be relevant in this case but once I know what is causing you problems I'll try to get it checked out

[BobWilliams757]

Just a little perspective from someone folding on a very basic system in terms of F@H.....

Many GPU projects (if not most lately) are painfully slow on this system. Often in terms of a day+ when folding 24/7, and not infrequently more than a day. BUT if they are assigned properly the only real difference in performance is the PPD spread that can only be accounted for within reason. Since many of us with slower gear aren't really racking up points, it's of no real concern to many of us. Whether the WU takes two hours or two days, it's accumulating some points that represent the contribution to science. And having gotten used to running a slow system, it's really of no concern to me as long as I'm meeting the timeouts. BUT it would be nice to have a way to opt out when it doesn't meet the timeout. I don't want duplicate work being done to get the science done, and especially if it makes me feel as if I've slowed the science down.

As for CPU projects, this thing is far from cutting edge but not a dog. Being that we use this system for other things, I tend to only CPU fold towards the end of the "other use" day. No sense in listening to the fans all day. But once I've picked up a WU I just let it run 24/7 until it's done. Some of these new larger WU's might allow people to just let it CPU fold for part of a day over multiple days, or fold on less threads. Only time will tell.

I do understand their are limitations to assignments, and that cherry picking can take place if there are too many options for us. The researchers can only do so much to keep everyone happy as far as preferences go. Maybe at some point of time there will be a system so most users more frequently get the WU's that suit their system efficiency and desires. But until then.... I still just fold them.

I have seen that some of these upcoming CPU WU's have generous points as well as timeouts. But they may well be more "bang for the buck" then some of the CPU units we are getting now. It all really just depends on the assignment restrictions. Until we pick one up, no sense in second guessing the possible outcomes.

[aetch]

It's 10 million steps.

Something with 2.5 or 5 million steps, a minimum core/thread count of 12 or 16, and a tighter timeout of 2.5 - 3 days I would've thought as more reasonable for something like this. You know, settings to ensure the science is done at a reasonable rate and FAH are not waiting a week to check up on the unit. I have no issue with getting a series of smaller units which ensures a smoother flow of science.

Never mind, I want to reconfigure my system and I'm not going to wait while this unit holds my system to ransom.

[BobWilliams757]

Are you already running one of these Aetch?

I agree it's without doubt a big CPU work unit, but with 7 days before the timeout it doesn't seem unreasonable to me. I've recently completed a CPU work unit with similar atom count and 2.5 million steps, and it was only slightly over half a day, while not even using full CPU resources. Assuming 18401 would take 4-5 times longer, it really isn't something that would be all that terrible to run.

As far as a system being held ransom, simply pause it and go about what you want to do. I've done it for updates and such in the past. There is no obligation to run 24/7 at full power available while folding.

While it would be great if we all got WU's a la carte per our systems, it's an imperfect system based on the research. I have to trust that the researchers involved wouldn't tie up our hardware for longer periods if there were options to make the science quicker while accomplishing the same end goal.

[JimF]

P 18401 is taking 32 hours on all 32 cores of my Ryzen 3950X (Ubuntu 20.04.3).
That gives it something to do.

[aetch]

Ok, cat among the pigeons time.

I'm curious, are you doing anything to lower the performance of your 3950x?
Are all of your threads being used for the work unit?
Are you getting full cpu utilization?

I had one of the 18401 units through my Ryzen 3900X, it took just shy of 40 hours.
My Ryzen is downclocked to 3.33GHz and the slot size was 21 threads.
I would've thought a full bore 3950x (~50% more threads) would've been a *sub 30 hour run.

[JimF]

Good questions. As I said, it is on all 32 (virtual) cores/threads; I am not running the GPU now due to the heat.
TOP shows 3178% CPU utilization.
And it is water cooled by a two-fan Arctic water cooler, and Tdie shows 63 C, so it is not down-clocking due to temperature.

I have used a 3900X occasionally on Folding too, and find that it is not as much slower as you would expect.
On other (BOINC) projects, it is probably due to the larger cache/core ratio, but I doubt that accounts for it on Folding. It is probably just inherent in how Folding cores work.

PS - The 3950X (and 3900X) are two CPU chips packaged together, and stitched together with some sort of matrix. That probably gets in the way.
The 5950X still includes two CPU chips, but improves on the matrix a bit. I would like to see it's numbers.

[Joe_H]

I would've thought a full bore 3950x (~50% more threads) would've been a *sub 30 hour run.

Depending on the number of atoms in the WU and geometry of the system, most projects will show a "knee" in the processing speed versus number of CPU threads. After a certain number of threads adding more doesn't improve the time as much. So 16 threads might process a WU in about half the time as 8, but 24 doesn't get close to one third.

[BobWilliams757]

Forum user Paragon did quite a bit of testing with various settings on the 3950X. I'm not sure how much of it would translate directly to the 3900X, but it does show the "knee" that Joe H. is speaking about in the post above, at least when running Windows 10.

This was a while back so I think most of the testing was with the A7 core, but notes on the increase in performance with the A8 core.

https://foldingforum.org/viewtopic.php?f=38&t=35286

[psaam0001]

I've got the majority of my CPU core's busy with one of these WU's right now... I'll accept that it will take at least another 1.6 days to run it for over 360K points.

Note I'm using 29 of 32 CPU threads.... The other 3 are for GPU tasks, as I'm trying to get as much out of my Kepler and GT series Pascal cards while I can.

I know I said in an earlier thread that I retired my Kepler's, but I put them back in my desktop systems to help make up for lost production with the PSU recall mess.

Paul

[aetch]

Err, you shouldn't need 1.6 days (38.4 hours) on a 3950x. Do you have your processor underclocked or something?

I've had 2 units for 18401 on my 3900x so far.
The first was a 38.5 hour run on 21 threads.
The second was a sub 37 hour run on 22 threads.
Each was 400K+ points.

[psaam0001]

I may need to check something, though I'm not over clocking it.

[Update: Forgot to remove a piece of cardboard between the CPU fan, and heat sink, when I originally installed it. D 'OH]

Paul

[DeeGee]

Ah, so this is some new BIG WU? I was wondering how I got a WU with estimated 20+ hours even on my 5950x. It's also my gaming computer, so usually I just click "Finish" and wait for the current unit to work itself out. I guess in this case it's not an option. I have limited F@H to 27 threads, but even with that it does cause some hitching in games.

[psaam0001]

I may need to check something, though I'm not over clocking it.

[Update: Forgot to remove a piece of cardboard between the CPU fan, and heat sink, when I originally installed it. D 'OH]

Additional update (9/30/21): I will be looking at upgrading the CPU heat sink and fan to a larger one from Noctua. As I went to their site (https://noctua.at) to see what their recommendation was for the 3950X CPU... Turns out I should have been using either the 120mm heat sink and fan combo (NH U12S), or the 140mm heat sink and fan combo (NH U14S). I'll move the one that is installed now (an NH D9L) to my Ryzen 3 system when I do that upgrade. Though, I have a box fan aimed on the open case running 24/7, until further notice.

Paul [Updated with the correct web address for Noctua's main web page on 10/24/2021]