Large Windows Runners

[hmaarrfk]

This fundable has not been submitted yet.

• This fundable has not been submitted yet.

Motivation

Infrastructure needed to help package large windows packages.

Expected result

A certain amount of large runners that can "opt in". Runners must not be exclusive to any one package, but rather for use by the conda-forge community as a whole.

Work items

• Draft a sample donation policy.
• Draft a sample set of requirements for the machine.
• Do we need a visual studio license?
• Execute on the donation of pool of windows runners.

References

• The donation of large GPU runners.

Skills / personnel

Infrastructure setup for Windows runners.

Estimated costs and/or resources needed (optional)

I've been building cuda packages quickly on 128 GB of RAM. This can probably be setup for approximately $2k USD per machine.

We would probably need something like 10 large machines.

So about $20k of donated hardware + 10% FTE for maintenance?

[h-vetinari]

Isn't prefix planning already on donate some runners for this purpose in #231?

I mean, not that it would hurt to distribute the load a bit of course! ;)

[hmaarrfk]

I'm leaving this here open as a way for people to put money and work where their intentions are.

[hmaarrfk]

For these funded projects, we need a proposal, a SOW, a budget, and a maintenance plance. Even writing this would be greatly helpful!

[ericdill]

@hmaarrfk do you have more info on the machine specs you're looking for? also is the $2k/20k a price per month or price per year?

[hmaarrfk]

My pricing was more on "home built" infrastructure.

We would need a place to store these machines, and a way to cool them.

But I'm fairly certain you can do it in an Asus (Intel) NUC with like 96GB of RAM.

So a 1 time purchase of about 20k for hardware infra, and then some software engineer to setup and maintain (not sure how things are costed for the GPU runners).

Should last like 4 years.

These are just "performant" machines, no GPUs. (i7 processor with lots of RAM).

• i7-14th Gen Asus Barebones NUC - $700 USD
• 96GB RAM - $250 USD
• Samsung P990Pro 2TB - $200 USD

Price per system: Approx: $1150

Rack + Powerstrip + switch = Approximate $1000

So for like 10 system, 12,500 + some engineer to maintain them.

[jaimergp]

xref to our "lessons learned" blog post: https://quansight.com/post/building-a-gpu-ci-service-for-conda-forge/. Might be useful while deciding how to do this.

For Windows, licensing might be something else to consider if they are publicly available machines.

[hmaarrfk]

The windows licensing is important.

I'm going off:

So hopefully that is OK! There might be more to the story too. But it is a consideration!

[jakirkham]

Cross-linking the .cirun PR adding Windows CI for the PyTorch feedstock: conda-forge/.cirun#14

[Tobias-Fischer]

I guess we can close here @hmaarrfk?

[hmaarrfk]

I'm not sure.

Let's see how the runners work over the next 6 months

[hmaarrfk]

Perhaps we can now move this to a call to action to just provide any kind of funding:
https://numfocus.org/donate-to-conda-forge