Folding Forum

Hi,
Quite new on the forum but folding since years

I've recently received long WU that my PC isn't able to finish before timeout.

This happen on an atom D525, yeah, old, but works great as a home server and TPF (time per frame) is about 1 hours 43 mins, ETA is over 6 days
I have an atom D2700, yeah, less old but old, but works great which is a bit faster, with a TPF of 59 mins 11 secs and ETA close to 4 days, but I think this won't be enough

My question is related to the assignment of WU according to the CPU capabilities, These PCs works 24/7 but are unable to end units before the deadline.

How to get smaller unit or longer deadline ? This is a waste of time for science and for my little hard working machines :p

I believe that CPU work units are assigned based upon CPU core/thread count, not CPU capability.
There's also the thing that many Atom processors have hyper-threading.
I'd suggest lowering the CPU cores/threads available to your CPU slot to 2.

No way to get smaller/shorter WUs as such ... changing cause might avoid the larger/longer ones ... can't remember what core count restrictions the AS has for the larger/longer ones (if in deed they have) but if say they are 4 core and above then setting a 3 core slot might work but iirc your cpus were relatively low power when new maybe 10+ years ago so just might not make it ... I'll see if it is possible (fairly sure it isn't) for these cpus to be treated like mobile phones which I believe only get certain projects (but I might be wrong)

https://ark.intel.com/content/www/us/en ... 0-ghz.html

https://ark.intel.com/content/www/us/en ... 3-ghz.html
these Atoms lack AVX or AVX2 instruction extensions, so use the older, less capable SSE2 instructions.

You could try a 'max-packet-size' parameter, and set the value to 'small', that actually controls the size of the download, not run time duration.

Oops

Ok, thanks for answering me seriously
I'll try some of your solutions ...

It is a sad fact that eventually cpus or gpus simply can't keep up with the needs of the FaH researchers.

FaH indicates a preference for wus to be completed within the Timeout by both adding a QRB to the points for those completed within this and releasing the wu for another folder to complete (hopefully quicker) if it is not met.

The Expiration Deadline is the point where any wu return is deemed of no value (hence the dumping).

Over time projects evolve to utilise more folding resource as the general availability of it become more as technology advances ... and this means that older less capable kit becomes less able to complete wus within the desired timeframes, eventually needing to be retired. Recently there has been an increase both in the available resource pool of newer more capable kit along with a requirement from the researchers to make use of this with larger, more complex modelling challenges.

This means that there are older resources that have begun to reach the point where they simple will no longer cope - and any adjustments such as the suggestions above are tbh simply prolonging the inevitable ... for gpus Fermi are now longer supported and on the near horizon Kepler may well be too ... for cpus older models, say 10 years plus, that were minimal processing power such as the atoms you run are now struggling to meet the needs of the current crop of projects - and I would guess actually struggle to meet timeout even on the smaller projects?

As I understand it (and I am just a volunteer here) a simple rule of thumb as to whether a cpu or gpu is worth folding with would be:

Any cpu or gpu that occasionally misses the Timeout but completes pretty much always within the expiration is still adding value most of the time.

Any cpu or gpu that mostly misses the Timeout even if not missing the Expiration often may be adding little value as the wus are getting assigned to other folders irrespective of whether they then complete within the Expiration. If the wus are regularly completed first by other folders then really the value is next to nothing.

Any cpu or gpu that regularly misses Expiration is simply increasing the folders heating (and power bill) whilst adding nothing to value to FaH and actually doing more harm than good as it is simply delaying the progress of the science.

I believe that some folders utilise their FaH retired kit for other distributed systems which are less time critical and architected in a different manner.

If you have a CPU with AVX and you get an assignment you cannot complete when running 24x7 please report it on the forum along with the thread count. The project owners are interested in COMPLETED WU as much as you are. They can easily limit assignments to low thread count CPUs.

That may mean you get no assignments but in your case, that won't let you get an assignment you can't complete. Hopefully there will be other assignments that you can complete. In some cases, they may adjust the deadlines.

In some cases that might work for SSE2-only CPUs, but I can't promise anything.

There are currently some project that have super long run times, but don't scale well to high thread counts. It's unfortunately an impossible situation to handle with current tools ...

What are considered a high thread count? Thanks.

Typically that might be thread counts higher than 8 or 12. The options used when creating Core_A8 obscure this a bit more than Core_A7 which would outright fail at some higher thread counts with a failure to decompose message. But it can be observed with smaller atom count projects where going from say 4 threads to 8 almost cuts the time in half, but going to 16 may not see much improvement at all, or even run slower

Thank you. I have 12 cores/24 threads and am also running 2 GPUs. What would you suggest as an efficient running thread count?

Thanks.

I would suggest a 20 thread cpu slot as a good starting point and play around with it until you're happy.

By default the FAHClient reserves a single thread/core for each GPU and it typically assigns the rest of the threads to the cpu slot.
With any system I personally recommend giving the operating system one or two threads to itself to breath.
On a 24 thread system with 2 GPUs that would give you a 20 thread slot.

On my own 24 thread system I'm currently running a single GPU with a 20 thread cpu slot. For quite a while I was running a 22 thread slot and I'm not really seeing a difference in overall performance.

To be honest I'm disappointed, I feel like if a humanitarian organization would refuse my 20$ because they want 30$ minimum...