There is a problem with the way that HPC based resources are funded currently. The problem lives in several federal rules and is way beyond my pay grade.
The problem lies in how grants are funded. Currently if a researcher requests hardware funds from the granting agency this is a capital request. Money for capital equipment is not subject to the 50% (or higher) tax most schools charge on new grant dollars, these are known as indirects. If they purchase an allocation on an HPC service, they pay for the service with after indirect monies.
Don't get me wrong, indirects serve a very valuable need and should exist. The problem is when they are attached to services, like an HPC service. Unless your institution has rolled some of the costs of the HPC service into indirects, (you lucky bastards) researchers pay the full price for 'services' with after indirect money.
When you purchase capital (nodes, switches) there are no indirects to be paid. In many cases institutions have not separated out costs that are covered vs not. A good example is collages providing rack rent or covering power costs for capital hardware. Or the inverse: not covering the rack rent or power for an HPC service.
Realistically this is backwards. If grant recipients are not paying indirects to cover provided power/racks/staff for capital (indirect free money) they should pay for it as a fee to their unit. If they are buying a service where the fee covers power/racks/staff they should not be paying indirects.
In the end, high cost items like super computers, and allocations on super computers should be as fair as possible between the two models. Administrations are hobbling researchers with added costs in both cases currently.
Now why I think schools should move away from purchasing capital for every grant and start providing allocations? That is a whole post for another day. For those who have ran condo clusters before, I think will know why.
I think requesting that indirect costs be waived for HPC allocations is a good idea. Or that those indirect costs be applied directly to HPC infrastructure. I think it's not hard to make a good case (as you have) on how this would actually save money.ReplyDelete
I think the breakdown is simple to prove the point of cost savings and transparency. I see a few groups: Hardware, Software, Staff, Power, Rent. You could get finer grained than that, but I don't see the reason to. Just always enjoy large institution politics.ReplyDelete