Folding on 3990X only using 30 cores

[Kompetenz Klaus]

Hi,

i want to fold on my TR 3990X workstation, while some cores are idle (happening a lot at the moment). However, i can only get 30 cores to work. Also, i only get like 3 WUs per day. Another problem i have is, that i don't get any WUs whatsoever for my two quadro fx 4800. (Have thrown them in for now, because i don't really do GPU intensive tasks, but GPU folding brings more points =D). 

[Spotty]

1 hour ago, Kompetenz Klaus said:

However, i can only get 30 cores to work.

There's a limit to how many threads it can use per slot. I think 32 but don't quote me on that.
Way around it is to add an additional slot. You can have multiple CPU slots and split the cores up over them. For example you could have 4 CPU slots with 30 threads per slot. Go to the Configuration Menu then in Slots add another CPU slot and assign the number of CPU cores (threads) you want to assign to it.

 

1 hour ago, Kompetenz Klaus said:

Another problem i have is, that i don't get any WUs whatsoever for my two quadro fx 4800.

I believe the Quadro FX4800 are going to be too old and aren't supported by F@H. The oldest supported GPU series for Nvidia is the GTX400 series (Fermi). The Quadro FX 4800's are that old that it is unlikely they would be able to complete a GPU work unit before timing out anyway.

[lexxR]

@Kompetenz Klaus 

 

(for site manager when doing @ can you make it so it first filters who is commenting in this Thread first,Not the whole of LTTS, this is the only site i know of that auto fills from the whole site when it only needs to be from this thread unless the name does not exist on here then it moves to site list)

 

set the FAHcontrol to have 4 CPU slots so it spawns 4 clients (assuming they are all using 30 threads each, you might want to limit it to 30 when adding the other 3 cpu slots and change the first one to 30 as well)

 

i am unsure if the FAH work client is NUMA aware and might get stuck on Group 0,you Might have to use somthing like Process Lasso to move the 2 other CPU Clients to the second 64 thread group or launch the 4 of them manually using 

start /NODE 0 application1.exe  

start /NODE 1 application2.exe

(i don't own a 3990x so unsure if the client needs the above)

 

the 3990x has 2 Virtual NUMA domains as max threads windows can handle is 64 threads before it has to split them into groups (if the client is NUMA aware you see FAH load on all threads and won't have to use process lasso or command line above)

 

windows 10 hides that the 3900x is actually using 2 groups (only noticeable when you see only first 64 threads get loaded up or you right click on a process and set affinity) not sure why ti does not show in task manager that its using 2 groups (like 2 NUMA or group)

 

someone else might be able to show more light on this subject

[Kompetenz Klaus]

Thanks for your replies,

i already tried to add another slot with 32 threads, but over the course of three days, only the first of the two slots got WUs. Might be due to the low availability of WUs. 

 

Does it really help to use all the Threads? At least in FEA and CFD analysis, using all the threads is just creating overhead, since when running on 64 cores, every core is used to its maximum anyways. Should be the same in FAH, or am i missing something?

 

Quote

windows 10 hides that the 3900x is actually using 2 groups (only noticeable when you see only first 64 threads get loaded up or you right click on a process and set affinity) not sure why ti does not show in task manager that its using 2 groups (like 2 NUMA or group)

It sort of does. Normally in taskmanager, two corresponding threads are placed nex to each other. With the 3990X, the first 64 Threads all belong to individual cores.

[Gorgon]

5 hours ago, Kompetenz Klaus said:

Thanks for your replies,

i already tried to add another slot with 32 threads, but over the course of three days, only the first of the two slots got WUs. Might be due to the low availability of WUs. ...

set option: client-type to advanced in Expert Tab of Configuration dialog.

 

A lot of the Work Units (WUs) being modeled are CPU only so we definitely can benefit from consuming all the threads on that monster. As a side benefit your get to stress-test your cooling

[gramzon]

I'm not very experienced in FAH but I have seen multiple mentions of there being less jobs for higher CPU core amounts, you may want to consider assigning a lower amount of cores per client like 16 or less.

[warmon6]

10 hours ago, Kompetenz Klaus said:

 Does it really help to use all the Threads? At least in FEA and CFD analysis, using all the threads is just creating overhead, since when running on 64 cores, every core is used to its maximum anyways. Should be the same in FAH, or am i missing something?

 

As far as im aware, it can help some but it really depends on the current Folding Core and the WU on those cores when it comes to SMT/Hyper-threading.

 

https://foldingforum.org/viewtopic.php?f=38&t=31357#p304984

 

Quote

The older CPU folding core - A4, gets more benefit from HT than the newer A7 core. As best as I can tell, A7 is better at saturating the arithmetic unit, esp. FPU, and leaves little excess capacity for additional threads from HT to provide much additional computing throughput. So running at a CPU thread setting that will use all available threads gives only a small boost over just using a number of threads equal to the number of cores. With A4 I would see a 15-20% boost using HT, much less on A7 WU's.

 

Basically what you are assuming is correct on the overhead. (Topic thread from F@H forum was about i5 vs i7 and hyper-threading but i dont think it'll be much of a different answer with current Ryzen/Threadripper SMT)

 

Quote

i already tried to add another slot with 32 threads, but over the course of three days, only the first of the two slots got WUs.

Which makes me wonder if F@H has added something to client to prevent users from using the extra spare threads in these high core count cpu's.... Wouldn't put it past them for doing such but i have no proof on that matter.  

 

Although it's probable what lexxR mention already is the more likely cause combined with the WU shortage. (Whether is a windows limitation or client, i couldn't tell you).

 

 

[Kompetenz Klaus]

Hi all,

it looks like my problem was due to a lack of WUs. I configured 1x32 and 2x16 slots, so i can manage my folding according to the workload, that is running on my PC for work. Yesterday i managed to get two WUs at the same time, so i assume it works now.

 

I have another problem though. A friend is running FAH on his laptop and is averaging 24k points per WU, while im only getting an average of 8k per WU across two systems. We have both entered the passkey, but he has entered at in the beginning, and i only requested it after 36 WUs where already done.

12 hours ago, Gorgon said:

A lot of the Work Units (WUs) being modeled are CPU only so we definitely can benefit from consuming all the threads on that monster. As a side benefit your get to stress-test your cooling

My cooling is already stresstested through FEA analysis. I have a beQuiet dark rock pro tr4 mounted, that on paper is only able to deliver 250W of cooling. However, when i run an analysis, all 64 cores are boosting over 4 GHz consistently. Really cool to look at, in taskmanager =D

 

12 hours ago, gramzon said:

I'm not very experienced in FAH but I have seen multiple mentions of there being less jobs for higher CPU core amounts, you may want to consider assigning a lower amount of cores per client like 16 or less.

Seems to work, thanks.

 

5 hours ago, warmon6 said:

https://foldingforum.org/viewtopic.php?f=38&t=31357#p304984

Interesting 

 

5 hours ago, warmon6 said:

Which makes me wonder if F@H has added something to client to prevent users from using the extra spare threads in these high core count cpu's.... Wouldn't put it past them for doing such but i have no proof on that matter.

I didn't even try to use all the threads, i just wanted to fold on 64 cores, so i had two slots with 32 cores. But as mentioned above, it seems to work now. Could be, that due to the time difference between Germany and the USA, there aren't that many WUs available, when i'm awake. 

[lexxR]

more work units seem to be around now (4x30 threads so it uses all 128 threads, i say 30 threads just in case of overhead but really up to you to expermant)

 

Make sure that it is actually using the second Group , all 120 threads should have load (as the 3900x is 2x64 on windows, on Linux unsure if it works that way) i recommend use of process lasso to see what its doing

[lexxR]

On 4/8/2020 at 5:24 PM, Kompetenz Klaus said:

Thanks for your replies,

i already tried to add another slot with 32 threads, but over the course of three days, only the first of the two slots got WUs. Might be due to the low availability of WUs. 

 

Does it really help to use all the Threads? At least in FEA and CFD analysis, using all the threads is just creating overhead, since when running on 64 cores, every core is used to its maximum anyways. Should be the same in FAH, or am i missing something?

 

It sort of does. Normally in taskmanager, two corresponding threads are placed nex to each other. With the 3990X, the first 64 Threads all belong to individual cores.

are you sure it does that it should be 32+32 cores (with each corresponding SMT thread per core making it 64+64 threads)

 

using a monitoring app per could would show if the cores are been used up first as the clock speed will be high on all 64 cores