Page 1 of 2

Old version download?

PostPosted: Sat May 09, 2020 2:03 am
by Kililea
Is there a place to get 7.5.1 until 7.6 is fixed? Hard to believe the speed issue hasn't been addressed yet.

Re: Old version download?

PostPosted: Sat May 09, 2020 2:11 am
by Joe_H
What speed issue are you writing about? The folding itself is handled by separate folding cores that are downloaded separately. Two of those have not changed, and the other was recently updated.

As fo older versions, there is a link to those at the bottom of the Beta version page - https://foldingathome.org/beta/

Re: Old version download?

PostPosted: Sat May 09, 2020 2:25 am
by JimboPalmer
There are rarely any advantages in folding on an older version of the client, which just handles I/O and UI, not science.

Core_a7 is the fastest CPU core ever. It does very well on 'new' CPUs, yet can fold all the way back to P4 Willamette.

Core_22 is the fastest GPU core ever. It works on any GPU that supports both OpenCL 1.2 and Double Precision floating point math.

Core _21 is a good, older core, supported to complete research started before Core_22 was written. It will not work on Navi (RDNA) GPUs and also needs OpenCL 1.2 and Double Precision floating point math.

You will not see much Core_21 work until COVID-19 work is done, as all that was started after Core_22 and so use the newer Core.

Some 'slowdowns' you may experience:
Windows on the Desktop has issues with more than 32 threads. Linux does not, nor does Windows Enterprise.
Some pricey graphics cards have so many shaders that smaller proteins may not use all the card to full utilization. This does not slow the Work Unit, but the card may have completed a more complex WU in the same time.
On the ides of March, about 10 times as many volunteers started turning in WUs than ever before. Here at the beginning of May, the servers are upgraded enough to keep issuing WUs on demand. The newer client will help you get more work than previous clients.

Re: Old version download?

PostPosted: Sat May 09, 2020 3:17 pm
by Kililea
On the same 7 core 4th gen machine 7.5 folds 2-3x faster than 7.6 resulting in at least double the PPD using the older version.

Re: Old version download?

PostPosted: Sat May 09, 2020 3:24 pm
by anandhanju
Unless you are seeing FAHClient 7.6 taking up a substantial amount of CPU % (there haven't been any reports of this happening), you should be seeing identical performance. As JimboPalmer's stated, the clients don't actually generate your PPD -- those are done by the science cores, specifically FAHCore_a7 for the CPU, and this has not had any new versions.

Can you post some representative PPD or time per frame numbers forthe two versions, along with their project numbers?

Re: Old version download?

PostPosted: Sat May 09, 2020 5:01 pm
by bruce
Kililea wrote:On the same 7 core 4th gen machine 7.5 folds 2-3x faster than 7.6 resulting in at least double the PPD using the older version.

That's simply not possible as has already been said. FAHClient spends most of the time doing absolutely nothing but waiting for you to type something in FAHControl or waiting for one of the FAHCores to finish working on a WU or waiting for the internetto give you more work to do. My guess is that it spends maybe 60 seconds per day doing something useful. So what if it actually can finish that 2-3x faster ... you might save 30 or 40 seconds per day.

You have been deceived by the initial estimates that are made when a new WU starts and there's zero information on which to base the initial projection. Once a WU has completed maybe 5%, you have provided it with a measurement of progress and the time it actually takes on your system so a rate can be estimated.

Re: Old version download?

PostPosted: Sat May 09, 2020 5:44 pm
by ajm
This morning, I swapped GPUs on my machines and a 1050ti landed in a PC running FAH 7.6.13. During over three minutes, at the first job allocated to that 1050ti, its PPD was around 2.2M (10-12 times its possible output). :shock: :D

They are right, Kililea, the version of the client doesn't affect the processing speed, at all. This software has nothing to do with it. But at the beginning of the folding, the estimated PPD can go all over the place.

Re: Old version download?

PostPosted: Sat May 09, 2020 6:30 pm
by Kililea
Yes the rate can vary wildly at start (in fact 7.5 takes several minutes longer to actually ramp up the CPU than 7.6), but it settles in to a pretty steady number during the hours the job is processing. That settled state is 2-3x higher for jobs processing under 7.5 and corresponds to the final credits on submit. I monitored this over 3 days after the upgrade because I saw other posts mentioning issues with 7.6. Immediately upon downgrading I see the previous rates again. The job times also are 2-3 times longer under the new version, correlating to the 2-3 times less PPD. So I appreciate you're saying this is not expected, but it's not only possible it's actually happening. Unless of course the numbers across the board are wrong or greatly misleading which bruce seems to imply, but then that makes all of this an unknown and all the numbers pretty meaningless which wouldn't make sense. Now I don't quite get what you're talking about with these "cores", I assume that has something to do with the code under the hood of the client version, but that seems irrelevant to make the distinction if they correspond to each other. Happy to help troubleshoot the issue by posting examples (if that actually helps), it will just take several days to collect that info of course.

Re: Old version download?

PostPosted: Sat May 09, 2020 6:42 pm
by Joe_H
You will have to post logs showing these difference you say you have observed. Quite likely they have another explanation than the one you have jumped to.

Re: Old version download?

PostPosted: Sat May 09, 2020 7:08 pm
by Neil-B
Cores are the functional code that actually do the folding - they are the "things" that actually do the folding and earn "points" - that is why everyone is saying that what you assert in the way of performance difference between the two clients must have a different explanation ... The Client does not do any folding - that is done by the cores - since they have not changed there should be no difference in folding performance.

It is possible that changes in setups between the two client installations could have caused such differences in performance - but these will have been because of configuration differences not because they are different clients.

Re: Old version download?

PostPosted: Sat May 09, 2020 9:19 pm
by PantherX
In addition to what has been stated above, PPD do vary by Projects so it could be a coincidence that you were getting Projects with higher PPD on V7.5.1 and then Projects with lower PPD on V7.6.13 but the variation is generally less than 10% and no where close to the 2x report unless you were using the estimate data from FAHControl before the actual data was collected which takes about 3% of the WU to be completed (bruce's point).

Re: Old version download?

PostPosted: Sun May 10, 2020 6:30 pm
by Kililea
Ok so Job running on 7.5 finished up with a PPD in the low 30k as normal. I upgrade to 7.6 and start a new job. Let it run for some time checking on it, take note around 4% and 6% and it's holding steady (and even dropped slightly) to just over 10k PPD as was the case over days previously running this version. Pause the job, uninstall (leaving data), install 7.5 again and start folding. Give it a while longer and once again it's holding steady over 30k PPD, and the time to finish is about half as well, and the TPF is also over 2x higher, all of which correlates to a higher PPD.

So, same job, so that's not the issue. No other system changes, just the client version. This is representative of the difference I have seen on two computers across the last week of observation. New client somehow makes slow folding. Yes FahCore is listed as 0xa7 in both versions of client, but clearly something is different. That or as I suggested 7.6 is just dead wrong in what it's showing so it makes it appear it's much slower than 7.5. I haven't actually timed an entire job so I suppose they could be finishing in half the time and giving 3x the points. I can try to catch a few and see if that's the case, but I notice even on submit the log says "estimate" so I'm really starting to wonder if any of this is remotely accurate. For that matter maybe 7.5 is over estimating and I'm getting less points than it indicates.

Re: Old version download?

PostPosted: Sun May 10, 2020 6:36 pm
by Neil-B
Please could you post logs as requested ... without these people will be unable to help sort out why you are seeing this effect

Re: Old version download?

PostPosted: Sun May 10, 2020 6:59 pm
by bruce
Estimates are based in the same information. The first time you get a particular WU, there is zero information so it's a complete guess. After a few checkpoints, it uses whatever data is has.

If you're running the same FAHCore on the same hardware with the same client settings, performance will not change.

The FAHClient version will not alter any of that unless you end up with different client settings.

When you download a new WU, it either uses the history you have previously generated or it starts with another SWAG until it collects some actual data.

Re: Old version download?

PostPosted: Sun May 10, 2020 11:48 pm
by Kililea
Apparently I'm not allowed to post files (while being asked to) and the gaslighting approach in these forums is tiring. I give up, too much effort while being told I'm not seeing what I'm seeing. I'll keep going by what the client says is happening on the off chance it's accurate and assume if it's not then it doesn't matter anyway.