HPC Software Stack Installation is the process of building the "Middleware" that sits between the Operating System and the user's scientific application.

In a normal server, you might just run apt-get install python. In HPC, this is forbidden.

Why? Because the generic version of Python from the internet is compiled for "generic" CPUs. It doesn't know how to use the advanced vector instructions (AVX-512) of your multimillion-dollar supercomputer. To get speed, you must compile everything from source specific to your hardware.

Here is the detailed breakdown of the "Pyramid" architecture, the concept of Environment Modules, and the automation tools, followed by the downloadable Word file.

1. The Software Pyramid

You build the stack in a strict order. If you change the bottom, the top falls.

1. Compilers (The Foundation):
• You need C, C++, and Fortran compilers.
• GCC: The open-source standard.
• Intel (OneAPI) / NVIDIA (HPC SDK): Proprietary compilers that often generate faster code for their specific hardware.
2. MPI Libraries (The Network Glue):
• This allows code to talk over the InfiniBand network.
• Examples: OpenMPI, MVAPICH2, Intel MPI.
3. Math Libraries (The Accelerators):
• Pre-written, highly optimized math functions (Linear Algebra, FFT).
• Examples: BLAS, LAPACK, FFTW, MKL.
4. Applications:
• The actual science code (e.g., GROMACS, ANSYS, TensorFlow).

---

2. The Solution: Environment Modules (Lmod)

HPC faces a unique problem: "Dependency Hell."

• User A needs Python 3.6 and GCC 7.
• User B needs Python 3.9 and GCC 11.
• You cannot install both into /usr/bin/python at the same time.

The Solution: You install everything into separate folders (e.g., /apps/python/3.6 and /apps/python/3.9) and use a tool called Lmod (Lua Modules).

• When User A types module load python/3.6, the system temporarily changes their PATH to point to version 3.6.
• When User B types module load python/3.9, it points to 3.9.
• They can switch instantly without conflicting with each other.

---

3. The Automation: Spack & EasyBuild

Compiling thousands of scientific packages by hand is impossible. (GROMACS depends on FFTW, which depends on MPI, which depends on a specific Compiler...).

HPC admins use Scientific Package Managers to automate this:

• Spack: The modern standard (developed by Lawrence Livermore National Lab). You can say: spack install gromacs %gcc@11 and it will download, configure, and compile GROMACS and all its dependencies specifically for GCC 11.
• EasyBuild: Popular in Europe. Uses "EasyConfig" files to reproduce exact software stacks across different clusters.

---

4. The Modern Twist: Containers (Apptainer)

Sometimes, compiling from source is too hard or the software is too old.

• Docker: Cannot be used in HPC because it requires "root" (admin) access, which is a security risk.
• Apptainer (formerly Singularity): The HPC version of Docker. It allows users to package their entire software stack into a single file (image.sif) on their laptop, upload it to the supercomputer, and run it safely without needing to install anything on the host.

---

5. Key Applications & Software

Category	Software	Usage
Module System	Lmod	The tool that allows users to load/unload software dynamically (module load).
Package Manager	Spack	Automates compiling complex scientific dependency trees (Python-based).
	EasyBuild	Automates building software stacks (popular in European HPC).
Compilers	GCC	The base open-source compiler collection.
	Intel OneAPI	Highly optimized for Intel CPUs (Xeon).
	AOCC	Optimized for AMD EPYC processors.
Containers	Apptainer	Secure container runtime for HPC (replaces Docker).