Intro to Nuvolos at ScPoEcon

Why Nuvolos makes Sense for an Econ Department

SciencesPo Paris, RES Data Editor

14 June, 2024

What is the Issue?

Modern (Economic) Research requires powerful computational infrastructure.
Many Machine Learning applications require access to GPU compute.
Large memory applications: Models with a large state space.
Bootstrapping of statistical models.
The requirement for powerful compute is no longer the domain of a few obscure macro economists (😈) who run their fortran via MPI on the HPC (High Performance Computing) cluster of their university. (Also: access to HPC clusters is still a topic for job market candidates each year.)
👉 Powerful compute requirements are becoming normal.

The traditional solution would have been to invest and maintain a dedicated HPC cluster.
Costly. Not long-lived (need to upgrade frequently).
Not overly user-friendly.
HPC cluster might make sense for large Universities with dedicated staff (and CS research departments…)

Low and Behold

SciencesPo does not have any high performance computing infrastructure for that reason.

The world has moved on: hyper scalers like AWS, Azure etc provide elastic compute services on demand, and at relatively low cost.

Cheapest available compute.
For a small org like an Econ Department (or a single researcher) the setup costs are prohibitive.
Not user friendly. Demo here (probably outdated).

Sits on top of Azure infrastructure: access to virtually unlimited computing power.
Offers 24/7 access to computing instances. Their business model is to figure out competitive pricing for provisioning those machines…
… and to provide world-class support.
Much more user-friendly than hyper scalers.
Built for small research teams and teaching which requires compute.

You can run any application on nuvolos. Behaves like a standard Linux machine.
(You can not run applications which run on windows only for the the same reason. 🤷🏻‍♂️)
For licensed products you need to bring your license; either as a file or as a network license.
STATA, matlab, gurobi, INTEL MKL, knitro, mathematica etc
Of course all open source software is available.
Support on the platform is 5 stars. 🥇

Small Research Teams: Your RA works on the same machine as you.
Individual Researcher: Running a 48h job on nuvolos is more (energy) efficient than running it on your office computer.
Teaching classes with significant compute setup costs or high demands (like machine learning etc)