Anandtech vs Tom's Hardware Folding@Home Coronavirus Race thread

[Endgame124]

No don't worry about it mate, it's not down to you who gets the WUs, just luck of the draw, but if you're having a good run of luck, feel free to spread it around .

Ref GPUs, ah yea, as per earlier posts apparently Nvidia GPUs need more CPU time, for F@H at least. I thought I could get away with BOINC using all my CPU & run GPU F@H without losing ppd, I did for a while but it seems that doesn't work with all F@H WUs . So I've left a core free again now.

I've been tracking my WU PPD in F@H on my A10-5800K running a 1660 Super. Turns out, on that ancient CPU, just leaving 1 core free was costing me around 150k PPD. I freed up a second CPU and things are running faster, but given we still have the occasional WU in F@H, I'll probably readd the extra rosetta thread and live with the hit to F@H.

 

[Markfw]

Find the folder in which the work unit is stored, and delete it.

For Windows it is

C:\Users\your user name\AppData\Roaming\FAHClient\work

.

You will see a few files, and some folders such as 00, 01, 02 etc. These folders are the work units. Look at your Folding client/interface to see what ID numbers are active, and delete the inactive folder(s). In the example below, I would delete the folder named 00

View attachment 20298

and for linux ? Most of my boxes are linux

 

[mopardude87]

No don't worry about it mate, it's not down to you who gets the WUs, just luck of the draw, but if you're having a good run of luck, feel free to spread it around .

Ref GPUs, ah yea, as per earlier posts apparently Nvidia GPUs need more CPU time, for F@H at least. I thought I could get away with BOINC using all my CPU & run GPU F@H without losing ppd, I did for a while but it seems that doesn't work with all F@H WUs . So I've left a core free again now.

Yeah still couldn't help but feel bad but your right

Yeah idk its odd sometimes i get like a 1k ppd increase letting the 7700k run then sometimes i lose ppd. I guess it depends on the wu? Only loads up 15-19% ballpark cpu usage letting the client do its thing only on the gpu. I get over 52k ppd on the 7700k by itself,pitiful compared to the estimated 1.43 million on the 1080ti. Seen some wu hit closer to 1.67 million. Some get closer to 1.2 million.

Why i got no issues giving the gpu as much cpu as possible as often as possible, no placement for displacement. Not even sure how the system will handle things with a 3700x and memory about 1ghz faster then what i currently run on this 7700k. Perhaps i could fold on both without a ppd penalty? Forgot that if i upgrade my 1080ti which i hope to do soon, i will gift it to friend and i will get her 1050ti and that will be my new primary. 3700x will be plenty for old games/folding and at night i will need to figure out how to get both 2080ti/1050ti gpus folding. I may hold off on a 2080ti and wait for Ampere but idk yet. I would prob splurge more on a 3900x if i end up playing the waiting game. Prob would be more inclined to do a 3900x/1070 primary/850w psu upgrade then hold off all the funds for the Ampere perhaps? 1070 is pretty boss i used a friends before, it played things fine. The 1080ti is overkill for 1080p anyways, lol thats why it folds more then games. Hasn't games in weeks since i started folding though.

I think to add a secondary gpu, i need to add in a slot into the client then designate its place as primary over secondary with a 0?

 

[TennesseeTony]

and for linux ? Most of my boxes are linux

File System / var / lib / fahclient / work

 

[StefanR5R]

Not even sure how the system will handle things with a 3700x and memory about 1ghz faster then what i currently run on this 7700k. Perhaps i could fold on both without a ppd penalty?

While I don't have such a system, I bet you still need to subtract 1 or 2 hardware threads from the CPU slot per each GPU for good performance.

I think to add a secondary gpu, i need to add in a slot into the client then designate its place as primary over secondary with a 0?

You just need to add another GPU slot, nothing more. There is no need to configure indexes.

I now have 10 units that can't seem to be sent.

Find the folder in which the work unit is stored, and delete it.

During the race, every WU uploaded eventually IME (and gave poor points then). I don't know how it is now.

 

[Assimilator1]

Perhaps i could fold on both without a ppd penalty?

Yea I reckon so, but as Stefan mentioned you'd need to leave a core free/GPU.

I wonder if some F@H WUs are like Einstein@H GPU WUs, those need a portion of the WU to be done by the CPU (can't remember why though!).

 

[Markfw]

While I don't have such a system, I bet you still need to subtract 1 or 2 hardware threads from the CPU slot per each GPU for good performance.


You just need to add another GPU slot, nothing more. There is no need to configure indexes.



During the race, every WU uploaded eventually IME (and gave poor points then). I don't know how it is now.

Its been over a week, and still they won't upload.

 

[StefanR5R]

Are the work server and collection server of these WUs shown to be down at serverstats? (At this moment, only one GRO_A7 server is down, i.e. a work server for CPU jobs. All other work servers are up.)

 

[biodoc]

There are lots of posts/threads at the folding forum regarding upload problems and apparently they've been trying to sort it out.

 

[mopardude87]

While I don't have such a system, I bet you still need to subtract 1 or 2 hardware threads from the CPU slot per each GPU for good performance.

You just need to add another GPU slot, nothing more. There is no need to configure indexes.

Yeah i forgot another gentleman from another thread mentioned the exactly affinity he used on his 3900x for folding. He did it by ccx cluster i think. Pretty much suggested if i wanted a gamer/folder i would be best off perhaps letting the first 12 threads game leave the affinity of the other 12 to fold. Might have been the other way around,i would need to ask again unless someone chimes in. I assume the 3700x has only 2 ccx clusters? Yeah still wrapping my head around how "ccx clusters" work but yeah first time taking folding seriously since my gtx580?

Been torn between 3700x and 3900x depending if i opt for a 2080ti or not, i was all in till the recent Nvidia May announcement. I heard it got pushed till October but recent announcement may HOPEFULLY mean a summer release like i figured it should have been all along prior the thing that should not be mentioned. Can't afford a 3900x and a 2080ti but a 3700x/2080ti is quite doable especially if i budget for a cheap B450 mobo. Meh been back and fourth right till this announcement. Almost glad perhaps as choices are becoming a bit clearer.

 

[IEC]

Yeah i forgot another gentlemen from another thread mentioned the exactly affinity he used on his 3900x for folding. He did it by ccx cluster i think. Pretty much suggested if i wanted a gamer/folder i would be best off perhaps letting the first 12 threads game leave the affinity of the other 12 to fold. Might have been the other way around,i would need to ask again unless someone chimes in. I assume the 3700x has only 2 ccx clusters? Yeah still wrapping my head around how "ccx clusters" work but yeah first time taking folding seriously since my gtx580?

Been torn between 3700x and 3900x depending if i opt for a 2080ti or not, i was all in till the recent Nvidia May announcement. I heard it got pushed till October but recent announcement may HOPEFULLY mean a summer release like i figured it should have been all along prior the thing that should not be mentioned. Can't afford a 3900x and a 2080ti but a 3700x/2080ti is quite doable especially if i budget for a cheap B450 mobo. Meh been back and fourth right till this announcement. Almost glad perhaps as choices are becoming a bit clearer.

I'll chime in here:
On my 3950X + nV system, F@H by default reserved 30 threads for CPU and kept 2 threads to feed the GPU. YMMV

 

[mopardude87]

I'll chime in here:
On my 3950X + nV system, F@H by default reserved 30 threads for CPU and kept 2 threads to feed the GPU. YMMV

Oh wow, out of the box set up it did that for you? Not bad assuming the 3950x unlike my 7700k does not tank ppd which i assume it does not? When i am not gaming i wouldn't mind like 22 of the 24 threads running on the 3900x assuming that is what offers the cpu/gpu the best overall performance. I would think it would default to the best overall performance where you shouldn't really need to tinker with it.

I just heard about the ccx cluster tweak and thought it was worth mentioning, the 3950x is in my price range honestly but i heard for gaming and stuff that chip isn't exactly boosting to the clocks of the 3900x. I may pick it still especially if i opt out of the 2080ti which is looking very likely now.

Hows it running for you? Not losing gpu ppd i assume but whats that chip doing for you on ppd? I only lose ppd or i gain 1k ppd or so as mentioned before when i let the 7700k do its thing but the cpu only loads upwards of like 86%? I last ran this before i enabled HD630 onboard as primary so perhaps shared bandwidth between IGPU and the cores may be worst off now? Thanks for your response btw

 

[IEC]

@mopardude87

Yes I don't recall having to set threads. It just reserved 30 for the CPU.

As far as PPD, CPU generally does not contribute much (can be as poor as 30K ppd with the "wrong" WUs). However, at least during the last race I had a stretch of a couple of days where I was getting mostly beta workunits that yielded me 470K+ ppd.

 

[mopardude87]

@mopardude87

Yes I don't recall having to set threads. It just reserved 30 for the CPU.

As far as PPD, CPU generally does not contribute much (can be as poor as 30K ppd with the "wrong" WUs). However, at least during the last race I had a stretch of a couple of days where I was getting mostly beta workunits that yielded me 470K+ ppd.

Not bad i was curious as i keep seeing next to nothing or a ppd lost on the 7700k its why its not even folding. Prob needing a ton more cores, faster memory and various other things to not have to put up with such non sense.I tried the affinity thing and it did nothing but still hurt the ppd of the 1080ti . 470k ppd wipes the floor with the 52k+ the 7700k gets. Your 3950x gets what my friends 3gb 1060 gets essentially at times.

Still more likely to get the 3900x of course, has better binning on the boost for gaming and stuff from what i heard and understand. Took much to not consider a 9900k as i am heavily cpu bottlenecked in mostly everything at 1080p in games. Not paying all that money for 8 hot ass cores on a dated platform LOL. I think Ryzen will come out on top with next gen games anyways i think.

 

[StefanR5R]

When i am not gaming i wouldn't mind like 22 of the 24 threads running on the 3900x assuming that is what offers the cpu/gpu the best overall performance.

Yes, that should be best for PPD as long as it works. But 22 is factored into 11 x 2, which could perhaps lead to WUs failing to start due to domain decomposition failure, requiring manual intervention. (Either dump such a WU, or set the slot to a lower thread count which is a multiple of small factors. While you aren't in front of the computer to watch out for such failures, they can block the CPU slot for a long time. My 88 = 11 x 2^3 threaded slots worked for a while, during the race, but then I had to give up on them and cripple the computers with smaller slots because of too frequent decomposition failures with 88 threads.)

I just heard about the ccx cluster tweak and thought it was worth mentioning,

I can't comment on manipulations of CPU affinity since I don't use Windows anymore on any on my Folding/ BOINC computers. (While it is possible to play with user-defined CPU affinity on Linux, there is no practical reason for it.) But I'd say, if you run another somewhat CPU intensive load besides Folding@home, then simply pause the CPU slot for this period.

i keep seeing next to nothing or a ppd lost on the 7700k its why its not even folding. Prob needing a ton more cores, faster memory and various other things to not have to put up with such non sense.

I have 7700k + 2x 1080Ti myself,* and these work perfectly well for Folding@home (with GPU folding slots only). That's with Linux of course, but Windows should work OK on them too, just with somewhat lower average PPD because of much more copying between VRAM and RAM on Windows.

For reasonable PPD from CPU folding slots, you need tons of cores.

--------
*) three computers of this config, purpose-built for GPGPU computing

 

[mopardude87]

For reasonable PPD from CPU folding slots, you need tons of cores.

Mainly want the cores as i want dual gpus in this rig, one doing nothing but folding mostly besides a new/next gen title load up on rare occasions and the main just a desktop driver/older title gamer/secondary folder .Hoping eventually the primary is a 1070 perhaps and i could game as well as fold. I do little gaming on HD630 but i want something of a dual purpose system. Perhaps i could work the 3900x in for folding too assuming it plays nice to the idea with 2 gpus folding in the late evenings when i sleep

I prob will find out how much of a performance hit can be had if like for example BF5 was running on the 3900x/1070 and i had the 1080ti just folding. If the ppd penalty was like 10% it may be worth but if this 7700 only gets up like 19% cpu usage with gpu folding and BF5 barely loads up like 45% of a 3900x from what i seen in videos then in theory i should be ok .

Kind of a new build, a bit different but that's the end goal with it. I love to game and currently fold but if i could do both that would be great. Rather do it with one system cause the only active spare i got near me now being used is a i3 2100/gtx660 build. That build needs the 660 or else that thing would be in here. I may bench the 660 in here to see what sorta ppd lost i can expect with a dual gpu purpose build but the 7700 is considerable weaker then a 3900x too.

 

[StefanR5R]

For a Folding computer with two NVidia GPUs, a dual-core CPU with Hyperthreading or a quad-core CPU without HT would be plenty. Clock speed of the CPU should preferably be 3 GHz or more, although I had good success with a lower-clocked Xeon E5v4 on Linux myself (driving three 1080Ti).

Each FahCore_21 or _22 process cannot use more than one hardware thread of the processor. Additional hardware threads or cores are good to have for background work of the OS and the desktop UI. Add more cores for other tasks that the computer is planned to accomplish while folding.

 

[mopardude87]

For a Folding computer with two NVidia GPUs, a dual-core CPU with Hyperthreading or a quad-core CPU without HT would be plenty. Clock speed of the CPU should preferably be 3 GHz or more, although I had good success with a lower-clocked Xeon E5v4 on Linux myself (driving three 1080Ti).

Each FahCore_21 or _22 process cannot use more than one hardware thread of the processor. Additional hardware threads or cores are good to have for background work of the OS and the desktop UI. Add more cores for other tasks that the computer is planned to accomplish while folding.

You have been a big help, so if i was running lets say a 1070 and 1080ti together, there should be 2 FahCore_21-22 processors each using a single hardware thread? Mostly pushing core 2 on a FahCore_22 process but i see light usage on other threads but i assume that's cause of the 4 tabs one including youtube open in opera, would i be best to lock this FahCore process to to a affinity of core 2 or just leave it alone? Fair to assume off what ya posted its ONLY using core 2 anyways and my assumption on other threads is correct?

Given how little this thing uses, i guess my end goal in theory should be feasible then. I got no issues tinkering for best performance given the oddball goal i got here, but when i want triple the threads and double the gpu and double the ram with a 1ghz increase over my current 2666mhz kit i assume i could obtain my 2 main goals at least. Anything less then suddenly all that hardware seems VERY wasteful and i might as well cheap on a B450+3700X and save myself $400 ballpark over the 3900x config. Even if next gen titles push 16 threads, i got 8 spare for a FahCore_21-22 thread and for the os. Then when the cpu folds, there is the 22 or so threads left on the 3900x that can chime in too unless i need to optimize that given my dual gpu folding during bedtime hours goal.

Sorry for all the questions, still obviously new to the folding scene i just got these goals and clearly want the best overall performance for both main goals. Any help i get helps the team and i want maximum efficiency obviously. Any tips and help so far have been EXTREMELY helpful, thanks for that btw and the help so far.

 

[mopardude87]

Yup its official i placed the order last night for this mobo and ram to go with a 3900x! Hoping my stimulus comes today or soon then i will slap a NH-D15 or maybe a big AOE maybe on the 3900x then finish off the build with a 1tb MP600 nvme. So damn excited and hyped for these upgrades!

The motherboard

Are you a human?

www.newegg.com

The ram

Are you a human?

www.newegg.com

 

[StefanR5R]

so if i was running lets say a 1070 and 1080ti together, there should be 2 FahCore_21-22 processors each using a single hardware thread?

Correct. (While I used the term 'process' in #842, maybe a more common term for this, especially on Windows, is 'task'.) The FahCore_21/22 GPU application is a single-threaded program on the CPU; just its portion which runs on the GPU is massively parallel. In contrast, FahCore_a7 for CPU folding is multithreaded, of course.

Mostly pushing core 2 on a FahCore_22 process but i see light usage on other threads but i assume that's cause of the 4 tabs one including youtube open in opera, would i be best to lock this FahCore process to to a affinity of core 2 or just leave it alone?

You don't need to set an affinity for FahCore_22. If the Windows kernel shifts it to another logical CPU occasionally, there is no harm done. Rather, other workload which you have running besides F@h, e.g. a game, may benefit if the kernel is allowed to move the FahCore_22 task to another logical CPU when it sees fit.

Fair to assume off what ya posted its ONLY using core 2 anyways and my assumption on other threads is correct?

It sounds indeed like the kernel is preferring to keep it on the same core all the time.

For an I/O heavy application which I believe FahCore_21/22 to be, it is beneficial if the kernel keeps the thread on a core which is near to the device that is being accessed. This can be a concern on dual-socket computers with separate PCIe roots on each socket. The Zen1 based ThreadRippers and Epycs have multiple PCIe roots too, because they are internally organized like classic dual- or quad-socket computers. But with i7700K/ socket 1151, or Ryzen 3000/ AM4, each core has the very same path to each PCIe lane in the system, making this easy for the operating system.

For a computationally heavy application, it is beneficial to throughput if the kernel keeps each program thread on the same physical core for as long as possible. That way, the processor can make best use of its caches.

For a multithreaded application which is very sensitive to inter-thread communication latency (this may apply to some games), it is usually best if all threads of the application are concentrated on the same socket, or in case of Zen1 and Zen2 processors, on the same Core Complex. (The desktop variants of Zen1 and Zen2 based processors are divided into either 4-core/8-thread core complexes, or 3-core/6-thread core complexes. Ryzen 9 3900X has got four core complexes with 3c/6t each, AFAIK.)

I am not up to date how well recent Windows versions handle these different situations. But all of this is playing a very minor role on desktop computer hardware anyway; this gets more of a concern on larger servers.

 

[mopardude87]

Correct. (While I used the term 'process' in #842, maybe a more common term for this, especially on Windows, is 'task'.) The FahCore_21/22 GPU application is a single-threaded program on the CPU; just its portion which runs on the GPU is massively parallel. In contrast, FahCore_a7 for CPU folding is multithreaded, of course.

You don't need to set an affinity for FahCore_22. If the Windows kernel shifts it to another logical CPU occasionally, there is no harm done. Rather, other workload which you have running besides F@h, e.g. a game, may benefit if the kernel is allowed to move the FahCore_22 task to another logical CPU when it sees fit.

It sounds indeed like the kernel is preferring to keep it on the same core all the time.

Your pretty awesome for all the help Thanks!

I will prob leave the entire deal on a set and forget and if a situation comes up with issues i guess i could ask fellow folders what the deal is. I always seek any tips or suggestions that could help of course. Gonna be tripling my thread count, so yeah its gonna be a bit overwhelming and i heard there has been issues with AMD and their scheduler on apps but idk if this has been fixed or not?

I guess the latest chipset driver is worst cause of tanked performance? Trying to get ahold of the earlier ones if possible, any idea where to get them? Or best place i should say?

Edit: on behalf of another member with nearly same config i was suggested anything prior to october 28th AGESA update so i grabbed the october 25 driver from this list.

https://www.amd.com/en/support/previous-drivers/chipsets/amd-socket-am4/x570

 

[Assimilator1]

I guess the latest chipset driver is worst cause of tanked performance?.......

Damn! I didn't know that! Where did you read that? Got a good link about that?

 

[mopardude87]

Damn! I didn't know that! Where did you read that? Got a good link about that?

Yeah i saw it in one of the cpu threads here on Anandtech for Ryzen some mentioned it a bit ago, i remembered it and when i asked about it one of the members then made his suggestion which i just assume means there is prob merit to it. Usually you update things but i recall some complaining of performance degradation on the latest.

If the degradation is YMMV or not idk, i just remembered the complaints and made mention of them then got my suggestion from the member. I do know he is on a 3900x and a GB x570 as well, i forgot the exact model.

 

[Ajay]

Well, yesterday was my last day for this year. I didn't contribute much, but it was fun (lame number of GPU units). If I can screw up the cash, I'd love to have something in the basement buzzing along 24x7 like I used to.
Speaking of GPU WUs, this last one is stuck in limbo for the past 18 hours:

******************************* Date: 2020-05-01 *******************************
17:01:11:WU00:FS01:Sending unit results: id:00 state:SEND error:NO_ERROR project:14253 run:53 clone:0 gen:0 core:0x22 unit:0x00000000cedfaa925eab703d8c424bd7
17:01:11:WU00:FS01:Uploading 140.85MiB to 206.223.170.146
17:01:11:WU00:FS01:Connecting to 206.223.170.146:8080
17:01:11:WARNING:WU00:FS01:Exception: Failed to send results to work server: Transfer failed
17:01:11:WU00:FS01:Trying to send results to collection server
17:01:11:WU00:FS01:Uploading 140.85MiB to 52.224.109.74
17:01:11:WU00:FS01:Connecting to 52.224.109.74:8080
17:01:11:ERROR:WU00:FS01:Exception: Transfer failed

 

[Ajay]

Heh. 10 minutes after this post it uploaded. Weird.