Fold on a set schedule
Moderators: Site Moderators, FAHC Science Team
Fold on a set schedule
Hi. My power company is setting up Time-Of-Day meters to charge more during specific times of day. Is there any way to set a schedule on the clients to not fold between certain time of day boundaries?
-
BobWilliams757
- Posts: 534
- Joined: Fri Apr 03, 2020 2:22 pm
- Hardware configuration: ASRock X370M PRO4
Ryzen 2400G APU
16 GB DDR4-3200
MSI GTX 1660 Super Gaming X
Re: Fold on a set schedule
You can set timers with the LAR extension for V7. I'm not sure if the V8 version has timers or not, I haven't really checked it out.
Fold them if you get them!
Re: Fold on a set schedule
BobWilliams757, LAR for v7 looks like it is going to work... Thanks.
-
appepi
- Posts: 47
- Joined: Wed Mar 18, 2020 2:55 pm
- Hardware configuration: HP Z600 (5) HP Z800 (3) HP Z440 (3)
ASUS Turbo GTX 1060, 1070, 1080, RTX 2060 (3)
Dell GTX 1080 - Location: Sydney Australia
Re: Fold on a set schedule
A contributor gave me the same advice about LAR systems and I had sought it for the same reason: trying to minimise electricity costs by limiting folding to "off peak" times, which for me are 10pm-7am, while also trying to avoid delaying the science by pausing jobs. My supplier's rate is currently 119% higher for the "peak" periods (2pm-8pm) than for "off peak", while the "shoulder" rate at other times (8pm-10pm, 7am-2pm) is only 22% higher than "off peak". Also, on weekends there is no "peak" rate, and instead I get "shoulder" rates between 2pm-8pm. It works just fine, but there are some additional considerations. Economic Rule #1 is to totally avoid folding during "peak" times, for obvious reasons.
One snag with trying to schedule folding by time of day is that if you are running BOTH CPU and GPU jobs on the same device, they are not very likely to finish at the same time. So time-based scheduling really can't handle that situation at all, and GPU folding is much more electricity-efficient.
So Economic Rule #2 is to fold on GPU only.
I only do GPU folding here, since my "folding" devices are elderly HP Z-series workstations that don't have anything much to do other than stay updated and fold, and their GTX 10xx and RTX 2060 GPU's are much more electricity-efficient than their CPU's, even though most have 12 cores x 2 threads to play with. Since the GPU folding is also the only job on each device and they have similar overheads and settings I get predictable ETA's. After some experimentation I found that it was rare for jobs to run more than an hour or two beyond the time I set to "finish up and stop", so I generally use LAR to set them to "Finish up and Stop" at 7am, and see where they are at when I am up and about - which usually means letting them finish at shoulder rates, which don't cost that much more.
After they finish the device is of course on idle until I actually shut them down, which doesn't matter much for the Z440s with 2060s in them, because they don't consume much power on idle. However, the older Z600s and Z800s are still using a fair bit of power on idle, so I mostly don't use them any more. I also use GPU Tweak to limit GPU temperatures to 70 Degrees, which costs little in terms of reduced points - some other threads discuss that.
Economic Rule #3 is to check out how much power you are using on idle, since that is pure "waste", and adapt your schedules and devices accordingly.
Ideally, the assignment servers should know more about the receiving environment to which they are assigning work, but with so many unique environments - many of them "concurrent use" cases that are unpredictable - this is unlikely to be a cost-effective development. As it is, once a job has been assigned, the only available options for the receiving device are run, pause, or dump - and the last is more trouble than it is worth.
Week 1 of this evolving plan was in June 2022, and by week #52 it had added 1,192 Mpoints at an additional electricity cost of (AU) $405.52 or [16.6% of total electricity costs, at AU $0.34) per M points. [I note that LAR costs consumption at (presumably US) $0.10 per kWh and uses the peak consumption of a GPU when calculating the PPD efficiency of GPUs in the data it collects, but here I am calculating a total cost of all changes relative to past "careless" consumption, since it is silly just to focus on folding - it makes sense to schedule other electricity use too).
With further tweaking , by week #104 it had added 706 M points at an additional electricity cost of (AU) $204.96 or 9.4% of total electricity costs, at AU $0.29) per M points.
With further tweaking including a change to a better plan from the same energy provider, by week #136 (ie week ending 25 Jan 2025) it had added 710 Mpoints at an additional cost of (AU) $0.22, which is of course about zero% of total electricity cost and $0.00 per Mpoint. During this last period I have only been running the 10xx GPUs when they have to be on for MS updates anyway (and then starting at 10pm) or occasionally for testing something or other. I also tend to run the three Z440s with 2060s 24/7 on weekends since each day is 9 hours of off-peak and 15 hours of "Shoulder" at only 22% more.
So Economic Rule #4 is check out the energy plans on offer and pick the one that suits best.
One snag with trying to schedule folding by time of day is that if you are running BOTH CPU and GPU jobs on the same device, they are not very likely to finish at the same time. So time-based scheduling really can't handle that situation at all, and GPU folding is much more electricity-efficient.
So Economic Rule #2 is to fold on GPU only.
I only do GPU folding here, since my "folding" devices are elderly HP Z-series workstations that don't have anything much to do other than stay updated and fold, and their GTX 10xx and RTX 2060 GPU's are much more electricity-efficient than their CPU's, even though most have 12 cores x 2 threads to play with. Since the GPU folding is also the only job on each device and they have similar overheads and settings I get predictable ETA's. After some experimentation I found that it was rare for jobs to run more than an hour or two beyond the time I set to "finish up and stop", so I generally use LAR to set them to "Finish up and Stop" at 7am, and see where they are at when I am up and about - which usually means letting them finish at shoulder rates, which don't cost that much more.
After they finish the device is of course on idle until I actually shut them down, which doesn't matter much for the Z440s with 2060s in them, because they don't consume much power on idle. However, the older Z600s and Z800s are still using a fair bit of power on idle, so I mostly don't use them any more. I also use GPU Tweak to limit GPU temperatures to 70 Degrees, which costs little in terms of reduced points - some other threads discuss that.
Economic Rule #3 is to check out how much power you are using on idle, since that is pure "waste", and adapt your schedules and devices accordingly.
Ideally, the assignment servers should know more about the receiving environment to which they are assigning work, but with so many unique environments - many of them "concurrent use" cases that are unpredictable - this is unlikely to be a cost-effective development. As it is, once a job has been assigned, the only available options for the receiving device are run, pause, or dump - and the last is more trouble than it is worth.
Week 1 of this evolving plan was in June 2022, and by week #52 it had added 1,192 Mpoints at an additional electricity cost of (AU) $405.52 or [16.6% of total electricity costs, at AU $0.34) per M points. [I note that LAR costs consumption at (presumably US) $0.10 per kWh and uses the peak consumption of a GPU when calculating the PPD efficiency of GPUs in the data it collects, but here I am calculating a total cost of all changes relative to past "careless" consumption, since it is silly just to focus on folding - it makes sense to schedule other electricity use too).
With further tweaking , by week #104 it had added 706 M points at an additional electricity cost of (AU) $204.96 or 9.4% of total electricity costs, at AU $0.29) per M points.
With further tweaking including a change to a better plan from the same energy provider, by week #136 (ie week ending 25 Jan 2025) it had added 710 Mpoints at an additional cost of (AU) $0.22, which is of course about zero% of total electricity cost and $0.00 per Mpoint. During this last period I have only been running the 10xx GPUs when they have to be on for MS updates anyway (and then starting at 10pm) or occasionally for testing something or other. I also tend to run the three Z440s with 2060s 24/7 on weekends since each day is 9 hours of off-peak and 15 hours of "Shoulder" at only 22% more.
So Economic Rule #4 is check out the energy plans on offer and pick the one that suits best.