A few general and Texas-specific points: - You can buy different levels of support with different SLAs. Without paid support you are on your own. -
In terms of purchasing and billing aggregation, there are several
companies that will do that for you. You might check with your local AWS
rep (probably Sylvia Herrera-Alaniz.) - Texas DIS MAY have a
master purchasing agreement with AWS, which would be one approach to
get UTSA Purchasing to make it easier for you to purchase AWS services. -
If you have good centralization of purchasing for AWS, I2 and AWS
academic egress waiver programs will be relevant. If everyone uses their
own or departmental purchase cards then aggregation of usage is nearly
impossible and therefore no way to apply a waiver. We took a dive into AWS 12-18 months ago primarily for storage with a
few researchers and library electronic archives. Tech support and our
account rep have been responsive to our needs. We utilize the Texas DIR
agreement and the egress
waivers. We get one invoice with all the departments listed and their
associated costs making it easy monitor usage and chargeback.
======================= At Oregon State we have done the same. We contract direct with AWS, not through a reseller, but get the same egress waiver. We utilize the AWS structure of "Orgs" (Organizations) to roll up all costs from our individual AWS accounts in to a single payer account. We then distribute charges to each of the Oregon State Univ fund indexes on a monthly basis. We front-end the AWS console with our SSO to not require separate username/passwords for AWS console access. We do not add in a tax to manage this service for researchers which has helped make it a popular option. ====================================== Use ScaleX
https://www.rescale.com and AWS for all of their HPC needs: to give me access to resources that were identical to theirs without
letting me in their private network. They
would create different instances for each HPC software instead of
fiddling with modules etc. It was very convenient and easy to use for
non-HPC experts ====================================
On spot pricing: from (our former HPC director) Amy Apon https://pdfs.semanticscholar.org/5956/5abe21878ff34f7dcf297fccc878d44a161d.pdf
Seems to make sense for large embarrassingly parallel problems.
For regular cloud pricing, there was a paper at Henry's ACIREF session last year
http://www.oscer.ou.edu/acirefvirtres2018_talk_cloudvonpremise_wells_20180807.pdf .
You could change the assumptions- power,space, staffing, etc., but the
conclusion is roughly that cloud makes economic sense for centers with
O(50) nodes and down, and
does not at O(350) nodes and up.
We find our (heavily utilized) core-hour cost is a lot cheaper than regular Amazon prices. |
Wiki >