Had to dump a WU - will I be penalised?

[stealthswifttlk]

So I hope this isn't a stupid question, but maybe someone could clarify this for me please.

I'm not new to folding, but have returned after a long stint away from the project (last folded back in 2009) and what with the changes (think I was perhaps on V5 or 6 prior) I figure this is probably the best section for my question.

Everything's been fine with my setup since I restarted just before Christmas. I do not have any dedicated rig, and am just using my main desktop machine and it's GPU in it's spare idle time. This machine is not powered up 24/7, especially as it is in my bedroom, although I have allowed my machine to fold up to 16 hours a day some days, though not every day. I fold on a XFX Radeon 280X as my GPU slot, with a CPU assigned to my 3770K. Yes, my hardware is a couple of years old now I know, so isn't 'up there' with the great stuff now coming out of the 1070s and 1080s.

Anyway, to get to the point. I was assigned a work unit to the CPU slot that had a 3hour 30min TPF and was estimated to finish in 15 and a half days. Base and estimated credit were 1023. I didn't mind being assigned this unit at all but realized that there was no way for me to realistically finish it within the deadline that was set for the work unit because I don't fold 24/7. If it had perhaps required just one or two overnights then I might have allowed it, but as it was saying 15 days to finish (based on full time folding) with I think it was an 18 day deadline, I thought this wasn't a realistic proposition for my circumstances.

I didn't like doing it, but once I'd quickly run the numbers in my head, I idled the CPU slot then deleted the CPU slot from FAHControl (advanced control) to keep the GPU folding. I then re-setup the CPU slot and it assigned me a new SMP WU. My thinking was that if I can't realistically do the long WU within the deadline, that it would be verified against my account that it had been deleted from my slot and be reassigned to someone else whose hardware would be able to complete it within the deadline set. Is this correct or was I wrong? Separately, would I be penalized in any way going forward for that course of action? To me, it was a last resort but I figure others may judge this differently.

Any advice/opinion would be appreciated, thanks for reading.

[foldy]

Does not answer your question but maybe in FAHControl it helps to set Configure=>Expert=>Extra client options: max-packet-size=small

[stealthswifttlk]

Thanks for that, I'll give it a try. I thought that might limit me further though as I may then be assigned more WU's for single-thread based on size and figured that'd be a waste as I have an SMP capable processor(?)

Internet bandwidth isn't an issue for me really. I've already set next-unit-percentage to 100 as I can usually download in under 5 seconds but a single average TPF on CPU is usually about a minute, on the GPU it's usually about 2 and a half to 3 minutes so the next unit was dropping points whilst waiting for the previous unit to finish. Uploading is usually under a minute to up to 3 or 4 for a GPU WU.

[JimboPalmer]

If more than 20% of your Work Units do not complete successfully, you will not get the Quick Return Bonus.

If you delete the slot, and then recreate it, the server is told you abandoned the WU. This speeds up F@H.

[Joe_H]

There are multiple reasons that you may have got those estimates. But your i7-3770K is more than capable of folding any CPU WU's currently available and completing them within the deadlines even with not folding all of the time.

The first is that there is an installer bug that sometimes causes the first CPU WU to be downloaded to be run on just a single core of the CPU. That will cause it to take longer than normal. Second, the first time the client downloads a WU from a specific project it has no historical data on how fast your CPU will process it. So until several percent is completed, the estimates of TPF, PPD and completion time will not be accurate.

Your CPU has 4 cores and 8 threads it can process on. Reserving one for GPU folding leaves you with effectively up to 6 it can use for CPU processing. If it was my machine I would just set it to 4 manually in the FAHControl Configure for the CPU slot. You can choose your own setting, best are multiples of 2 and/or 3.

[stealthswifttlk]

If more than 20% of your Work Units do not complete successfully, you will not get the Quick Return Bonus.

If you delete the slot, and then recreate it, the server is told you abandoned the WU. This speeds up F@H.

Thanks for that, that was the main part of my question. Figure that I did the right thing by this then (judging from the 'speeds up FAH' bit)

I thought Quick Return bonus was based on returning within Timeout, the faster the better for Pande group/Stanford? By my quick calculations, for all historic WUs I think I've returned over 95% of everything I've ever been assigned. Back in 2009 I got frustrated with the project for a while as I had folded multiple WUs and then had them sat on my machine for ages (Core2Duo back then, and an 8800GTS) as they wouldn't upload. I had filled multiple slots in the work queue and eventually it got to the point where I had WUs going past deadlines set. I knew I'd done the work and was well within time but when I was waiting for 4 days or more for something to upload, it was getting silly so I thought not worth my efforts anymore until the platform stabilized. At least that's the way I saw it at the time, very hit and miss for me. Everything has been fairly smooth for me since restarting last month.

There are multiple reasons that you may have got those estimates. But your i7-3770K is more than capable of folding any CPU WU's currently available and completing them within the deadlines even with not folding all of the time.

The first is that there is an installer bug that sometimes causes the first CPU WU to be downloaded to be run on just a single core of the CPU. That will cause it to take longer than normal. Second, the first time the client downloads a WU from a specific project it has no historical data on how fast your CPU will process it. So until several percent is completed, the estimates of TPF, PPD and completion time will not be accurate.

I've folded approximately 100 WU's since I restarted last month, and estimate that there's approximately a 3:1 to 4:1 ratio of WUs being done CPU vs GPU (CPU units are usually about 1hr 30 to 1hr 50, GPU's have been averaging about 4hrs 30 to 5hrs 25). I left the CPU WU running for a while, and the time estimation was correct based upon the percentage complete over time that I monitored it. I had checked processor affinities and core usage and it was crunching across all cores available to it (Yes, hyper threading is enabled on my rig and I know I have 4 physical and 4 virtual cores because of this) This is why I was a little surprised as this WU seemed to be so far out of the usual sort of range. Percentage complete was slowly creeping up in fractions of a single percentage over such a long time. I had no reason to not trust the TPF based on what I was actually seeing as it was folding.

I'm in the UK so I'm ahead of the US in timezones. One Monday morning when I started up FAH recently there was no WUs being pulled down for the CPU. I checked service status and figured there were no SMP units available for me at the time as the servers were listed redline for me at that time so must have been a busy weekend. I switched to assigning a single core and the system then pulled down a WU for me that took about 16 hours. I was fine with that as was what was available. The estimates and complete times were accurate. I rely less on PPD as an estimate in case I need to pause or reboot for whatever reason.

Your CPU has 4 cores and 8 threads it can process on. Reserving one for GPU folding leaves you with effectively up to 6 it can use for CPU processing. If it was my machine I would just set it to 4 manually in the FAHControl Configure for the CPU slot. You can choose your own setting, best are multiples of 2 and/or 3.

I usually run on full slider, have slots set at -1 (auto) but have pulled CPU slot to full 8 occasionally if I've had a GPU slot finish and I'm coming to the end of the day and just want to finish the current CPU without starting another 4hr+ GPU. The 22%(ish) speedup is of course beneficial to allow me to power down sooner. CPU thread count usually pulls to 6 from 7 (8th thread being utilized for the GPU) because of the whole 'avoid decomposition by a prime number' thing. Auto seems to work okay for me for the moment.

[bruce]

When you have a.single GPU, the -1 (auto) setting for the CPU is (N-1). Adjusting the slider does change the meaning of -1 to lower values.

There are no longer any WUs that will assign to odd, prime numbers like CPU:7 or CPU:11. This will eventually be fixed by the beta client by simply reducing the effective setting to some smaller multiple of 2 and/or 3.

[bruce]

When you do need to dump a WU, FAH tries to report it so that it can be reissued quickly rather than remaining idle until the deadline. Those reports minimize FAH's loss of production. If that report is lost, the WU will eventually time-out and be reissued (but later).In either case, it's counted against your completion ration.

[stealthswifttlk]

When you have a.single GPU, the -1 (auto) setting for the CPU is (N-1). Adjusting the slider does change the meaning of -1 to lower values.

I was always aware that for GPU folding, there needs to be a percentage of CPU taken in order to control the workflow being sent to the GPU. On my old C2D (no hyper-threading, E6600) I have a recollection that this was affinity 1 and 50% of my CPU. I upgraded to a Q9550 (again no hyper-threading) in early 2009 (started folding 2008) and this opened me up to SMP folding and reduced the GPU to 25% CPU utilization - manually set to affinity 3 for GPU leaving 0, 1 and 2 for the CPU knowing 0 would be a lesser percent-aged core due to OS overhead etc.

Since I've been folding again on my current 3770K, from the behavior I've seen at least, I assumed (I know, often dangerous) that the GPU didn't need a full core for it's work assignment as the V7 client with all settings at auto just configured CPU 7 and on activation then pulled back to 6 to avoid degradation. This is fine with me - my partial assumption was that it aided OS overhead as well as it leaves some allowance for OS/other apps.

There are no longer any WUs that will assign to odd, prime numbers like CPU:7 or CPU:11. This will eventually be fixed by the beta client by simply reducing the effective setting to some smaller multiple of 2 and/or 3.

So I figure this means that optimal setting is best to leave it 'as-is' on -1 (auto) and let the client decide? It's been working that way so far, even though it configured CPU:7 by itself and still pulls WUs on there usually without issue, it just pulls back to 6 (according to the log) on initialization.

When you do need to dump a WU, FAH tries to report it so that it can be reissued quickly rather than remaining idle until the deadline. Those reports minimize FAH's loss of production. If that report is lost, the WU will eventually time-out and be reissued (but later). In either case, it's counted against your completion ration.

Okay, good info there - thanks. Do you know if the penalties are the same for either scenario? If something assigned to you goes past deadline vs is reported dumped/reassigned? I don't feel great about dumping stuff, but still think it was better in this case than for it to potentially go past deadline and be wasted time/effort for it to be rejected as re-assigned anyway and thus potential duplication of workload.

As an aside, I tested the max-packet-size:small on CPU and GPU. So far, it's pulled the same WU type for CPU PRCG 13746 and GPU said that there were no WU's available. I reset it and the first WU to come down was a PRCG 11432 - estimated at 12 hours so yeah, that setting seems to work. The GPU WU is a little larger than I've usually had, but where I usually do 2 to 3 GPU WU per day at approx 4 1/2hrs each when I'm folding, it just means there's less network activity and a more definitive folding timeframe for me today I guess.

[bruce]

The penalty to you will be the same .... An expired WU has a greater penalty to the science than if the notification is sent promptly.

In each case, the GPU needs one of your CPU threads, Your C2D had 2 threads, one out of 2 is 50%. Your Q9550 had 4 threads, so 1 out of 4 is 25%. Your 3770K has 8 threads, so 1 out of 8 is 12.5%. After deducting that one thread, you went from 1 to 3 to 7 that could be allocated to CPU folding, but 7 (and larger prime numbers) don't work reliably.