HPCOffPremiseCosts

HPC Off-Premise Costs Document

Summary

Cost estimates for hosting the NJIT HPC infrastructure off-premise (in the cloud) for a three-year period are presented for Penguin Computing, and are in preparation for Amazon Web Services. Cost estimates for Azure are forthcoming.

Purpose of the report

This report is the start of a costs-benefits analysis of hosting some portion - as yet to be determined, and likely to be fluid - of NJIT's HPC infrastructure off-premise.

Report

• Three-year cost estimates, as of 10/31/2017
• Seven-year HPC expansion cost estimate
• A cost estimate for hosting the HPC cluster infrastructure at Azure for 3 years is expected by EOD 11/03/2017
• Cost estimates for other vendors - Google Cloud Platform, IBM Bluemix, Oracle Cloud - may be obtained later, on an as-needed basis.
• Big data hardware off-premise cost estimates are not included in the current phase; to be addressed later.

Proposal

Assuming that it is cost effective to move HPC off premise, ARCS recommends using the Penguin Computing On Demand (POD) service for the following reasons:

• POD is entirely hardware-based
• Penguin Computing provides much of all of the software needed by NJIT, including compilers, utilities, and scheduler. This is not the case for AWS or Azure
• POD includes the Luster parallel file system on the high-speed Omnipath node interconnect
• POD is flexible with respect to accommodating user needs, e.g., backups

A trial project(s), involving one or more researchers, may be useful in gathering data on the deployment and use of off-premise HPC.