Estimating computing resources needed for a GWAS?

Question

One of my dissertation papers is going to involve a GWAS. I have never actually run a GWAS before, and do not know how to estimate the computing resources I need for it. I asked someone on my committee who said "it's all guesswork" but I don't even know how to come up with a reasonable guess!

I need to put some estimate of how much cluster time I will need (XX cores for YY hours) in order to do a dissertation grant. I've tried Googling and searching literature for how to do an estimate of this, but I'm not finding much.

Details:

• GWAS, total n=350,000, 10-15% cases
• Already imputed by the time I get it, will have about 96 million variants
• Will be using Plink
• Cluster nodes have 24 cores, and jobs can use up to 4 gigs of RAM per core. If Plink can handle being spread across nodes, I can use up to 100 cores at once.
• Time is billed in core-hours, using 20 cores for one hour is equal to using one core for 20 hours
• We may want to do a few smaller GWAS where we restrict by age (eg. one of all cases and controls under 25, and one of all cases and controls who are 65+)
• Need to come up with an estimate of cluster time needed for each GWAS that we might run (XX cores for YY hours)

Can you please advise me on how to estimate something like this?

Answer 1

Answer from @llrs, converted from comment:

I found an article where they said that they could do GWAS in 6000 samples in 6 hours in a regular user computer. But I couldn't find any other explanation about how does it take with other kind of data. It wasn't with plink too and I think it was a bit outdated.

Answer from @christopher-chang, converted from comment:

Incidentally, you may want to use plink 2.0 rather than 1.9 for this. It automatically makes good use of all your cores; in contrast, you’d need to start ~24 copies of plink 1.9 to saturate a 24-core machine.