What happens when software is retired or starts to die?

[terjekv]

Nothing works forever. Even if its "just" due to changes in openssl or TLS, not even the compatibility layer will save us forever. What do we do when this time comes? When do we start to drop software, how do we drop it, and what can we do to potentially support our users afterwards?

[mboisson]

Some experience from Compute Canada.
Our original stack (StdEnv/2016.4) still works and is still used. We deployed StdEnv/2018.3 when we got AVX512 hardware. New clusters adopted 2018.3, but old remained with 2016.4 as default. We are just now preparing a transition from both of them to StdEnv/2020

2016.4 used Nix 16.09, GCC 5.4, Intel 2016, OpenMPI 2.1.3
2018.3 used Nix 16.09, GCC 7.3, Intel 2018.3, OpenMPI 3.1.2
2020 uses Gentoo 2020, GCC 9.3, Intel 2020, OpenMPI 4.0.x

Each version supports more than one compiler/mpi/mkl/cuda, but the "primary" compiler and OpenMPI are fixed.

We have some documentation on the various options :
https://docs.computecanada.ca/wiki/Standard_software_environments

For our transition to 2020, our plan is :

1. First (in the upcoming weeks) announce to users that 2020 is available, why they should switch, and how they can make it their own default
2. Set a deprecation message on 2016.4 and 2018.3, and tell users they should try 2020 (January 2021)
3. Switch our default to 2020 (April 2021)