Folding on Cloud instructions (3 mil ppd)

[Knish]

I've put instructions on github of my latest updates/pictures from Google GCP and MSFT Azure to get you a free boost to folding power. Pay attention to your remaining credits and shut it down before you reach 0 to avoid charges. If you follow the guide, GCP uses < $3/day so should last 99 ish days. With Azure you only have 30 days to use your credits before they expire (about $5.5/day per GPU). [Update: GCP now has a 90 day limit which would leave about $40 unused by the T4, so you can throw in an additional instance of an E2 16core cpu for 14 days to use that up]

Don't try to use any virtual phones or virtual cards for sign up, as bots use those too so you could get an acct suspension

https://github.com/gitHu6-newb/FoldingAtAltitude

[MeeLee]

On google cloud, you can use like 5 or 6 shared cores, the slower ones, and last an entire year with the $300 free credit.
Anything that is shared, will regularly be throttled down.
You will need to order a compute unit, to have it truly work 24/7.
That being said, most of the units on MS Azure, Google Cloud, or Amazon AWS are running in the vicinity of 2 to 2,5Ghz CPUs; and while the reported CPU speed is faster than my Atomic Pi (running only at 1,69Ghz), I found that a single Atomic Pi (running at 15W at the wall = ~$15/Yr on electricity) is for some reason running about 2x to 3x faster than the AWS services that report 100% CPU utilization in their chart (have not been throttled).
Which leads me to believe that they're running hyperthreading as a second core, and aren't as forthcoming with the requested performance (at least not on the cheaper models).

If you ever run out of credit and want to purchase a quadcore unit, you'll lose $150/Yr at least.
For the same price, you can buy 3 Atomic Pis on amazon, and outperform them by 3x in 1 year.
Run them 2 years, and 6-7 units will perform them by 6x.
They're older, slower technology, but for the price and power consumption, they outperform AMD Ryzen 9 3000 series CPUs!
Meaning, 18 units perform better, consume the same power, and cost the same as a Ryzen 9 3950x with motherboard, ram, PSU, and case.
(This is thanks to the cores being energy efficient Atom cores, running at half the frequency of the Ryzen)

[gunnarre]

A 2080 Ti rented on vast.ai costs about 0.2 USD per hour and gives about 4.3 million PPD if the CPU is also fast (faster feeding). It costs less if you bid on an interruptible instance, but I would not recommend that because it might be outbid and stopped until timeout. If you get the $2 free credit, that translates to about 10 hours of free compute time on a dedicated instance. Unfortunately, some of the vast.ai instances don't work in Folding At Home, only with Tensorflow, and you only find this out when you test it yourself.

IBM Cloud also has GPUs, but note that their $200 free credit does not apply towards per-hour GPU compute instances, only towards rental plans, so there was no free GPU folding to be had there.

[kbeeveer46]

I tried Azure and I got to the step "Req quota increase" and I wasn't able to request an increase because I got the error "Your free trial subscription isn't eligible for a quota increase. To request a quota increase, first upgrade to a Pay-As-You-Go subscription. Learn more"

From your previous comments in the guide it looks like I can skip that step and still keep going.

[Knish]

make sure to wait a few minutes and refresh. there's some delays where it takes a few minutes for Azure to fully recognize that you have upgraded. You must be upgraded before you can request increases as well as use the GPUs

[PantherX]

Here's a link on how to configure GPU instances on AWS: https://aws.amazon.com/blogs/compute/fo ... instances/

It would be nice if you can include something similar in your repo... almost a one-stop-shop for F@H in cloud

[kbeeveer46]

After running the command

Code: Select all

az vm create -n AzureFolding -g myCloudFolding --image debian --generate-ssh-keys --size Standard_NC6s_v2 --priority Spot --max-price 0.30 --storage-sku StandardSSD_LRS

I got this error

Code: Select all

Azure Error: InvalidTemplateDeployment
Message: The template deployment failed with error: 'The resource with id: '/subscriptions/e11f9f69-cdb2-4921-8904-8dfd8ef2b2f4/resourceGroups/myCloudFolding/providers/Microsoft.Compute/virtualMachines/AzureFolding' failed validation with message: 'The requested size for resource '/subscriptions/e11f9f69-cdb2-4921-8904-8dfd8ef2b2f4/resourceGroups/myCloudFolding/providers/Microsoft.Compute/virtualMachines/AzureFolding' is currently not available in location 'eastus' zones '' for subscription 'e11f9f69-cdb2-4921-8904-8dfd8ef2b2f4'. Please try another sizeor deploy to a different location or zones. See https://aka.ms/azureskunotavailable for details.'.'.

Do I have to go back to the previous step and pick something other than East US? I ran the command after receiving the quota increase approval email.

[Knish]

kbeeveer, unfortunately I can't say with 100% certainty as I'm not sure why it is showing unavailable to you. As far as I can tell, that size is still available in eastus when I did an availability list command. If you tried again later and it still isn't working, then it looks like for you the only other US region that might work is South Central...

but first, it might be worth it to try their twitter help: @AzureSupport if they can confirm if NC6sv2 should be available to you in eastus or not.

If they can't, or they ask you to make an MSDN forum post (which usually leads nowhere) I can only think of to try the southcentral location, so first check your quota for SouthCentral this time, and increase the Spot vCPU to at least 6 if necessary. Once that's approved, delete our failed eastus attempt (which takes a few minutes) then try a new one:

Code: Select all

az group delete -n myCloudFolding

az group create --location southcentralus --name myCloudFolding

az vm create -n AzureFolding -g myCloudFolding --image debian --generate-ssh-keys --size Standard_NC6s_v2 --priority Spot --max-price 0.302 --storage-sku StandardSSD_LRS

the price is a bit steeper unfortunately, and if you get a price error, you'll have to increase it accordingly.

From there should be as per the guide again. Stay on top of your credit spending tho. And if after a few days it looks like your VM spends too much time getting pre-empted, you'll have to increase the price again while the instance is stopped with

Code: Select all

az vm update -g southcentralus -n AzureFolding --max-price 0.xxxx

If even that doesn't work, you can try multiple size Standard_NC6 (at least 2, and for a limited time 3 or even 4).You'll have to create them one at a time with different names, like AzureFolding2, ..etc. They can all be part of the same group 'myCloudFolding.' They are cheapest on southcentralus so you can use price of 0.13 but once again you may have to req quota increases for how many to use. (I believe these are K80 GPUs. I tried the multiple GPU size before like the NC12 but couldn't get all GPUs folding, so I advise staying with the NC6)

[Knish]

Here's a link on how to configure GPU instances on AWS: https://aws.amazon.com/blogs/compute/fo ... instances/

It would be nice if you can include something similar in your repo... almost a one-stop-shop for F@H in cloud

The reason I don't include AWS is there's no free credits offer, so I just focused on the two that people could do for free.

[gunnarre]

How much PPD per USD would one expect on AWS? I suspect it wouldn't be cheaper than vast.ai.

[PantherX]

...The reason I don't include AWS is there's no free credits offer, so I just focused on the two that people could do for free.

Good point! I didn't know that about AWS... seeing that their competition does it, maybe they might join that too! I am aware that they have a free tier (https://aws.amazon.com/free/?all-free-t ... -order=asc) but not sure if anything in them can be used for F@H.

[MeeLee]

...The reason I don't include AWS is there's no free credits offer, so I just focused on the two that people could do for free.

Good point! I didn't know that about AWS... seeing that their competition does it, maybe they might join that too! I am aware that they have a free tier (https://aws.amazon.com/free/?all-free-t ... -order=asc) but not sure if anything in them can be used for F@H.

You can run CPU, on a pretty slow 2-2,5Ghz CPU core.
It's enough for a month of running time.
Then shut it down.

[Knish]

You can run CPU, on a pretty slow 2-2,5Ghz CPU core.

"slow" because you need about 9hrs of idle-ing run time to earn 1hr of actual cpu usage at its full capability, otherwise you're really looking at ~ 200 MHz sustained usage, so I would bet any CPU WUs done with this somebody else ends up finishing it for you the next day while this happlessly chugs on for the next 3 or 4 days or until expiration.

[Knish]

kbeeveer: I tested Azure from scratch this morning and was still able to make eastus NC6s_v2 instances, so I hope you got it all sorted out

[Knish]

Update: GCP now has a 90 day limit to use your free credits which would leave about $40 unused by the T4, so you can throw in an additional instance of an E2 16core cpu for 14 days to use that up