Cache-Backed Images (CBIs)
A Cache-Backed Image turns “install and configure” into “pull and go.”
A Cache-Backed Image (CBI) is a minimal container — just Spack and MPICH — that comes already pointed at the full E4S build cache. It doesn’t carry the software; it carries the means to fetch any of it, already built and tested, on demand. The result: any of the hundreds of libraries, tools, and applications in E4S becomes available in real time, in full or à la carte, without a source build.
Why It Matters
Scientific software is usually hard to install. A researcher who wants to try a new simulation code often spends hours — sometimes days — resolving compilers, libraries, and dependencies before running anything at all. That setup cost is a real barrier to entry, and it falls hardest on newcomers and on scientists whose expertise is in the science rather than the software.
A CBI removes that barrier. Because the image stays small and the build cache supplies everything else, an install that would normally take hours can complete in about a minute.
Zero to LAMMPS in Two Minutes
On an ordinary laptop:
| Step | Time |
|---|---|
Pull the E4S Cache-Backed Image (ecpe4s/e4s-spack-cpu) |
42 seconds |
| Install LAMMPS from the E4S build cache | 75 seconds |
| Total: a fully functional, parallel LAMMPS simulation | under 2 minutes |
That run used LAMMPS’s standard Lennard-Jones melt benchmark — 32,000 atoms across 4 MPI ranks and 4 OpenMP threads each — and it wasn’t a toy run: 17.5 million atom-timesteps per second, with no source build anywhere in the critical path.
Why This Matters for AI-for-Science
Running a simulation is the entry point for a great many AI-for-science workflows — generating the training data behind a surrogate model, for instance. When the simulation step takes two minutes instead of two days, it changes what researchers are willing to attempt. A CBI turns simulation from a multi-day setup project into something you can fold into an afternoon of experimentation: spin up the environment, generate training data, iterate, and tear it down, without ever compiling anything from source.
Consistent from Laptop to Leadership-Class Systems
The same CBI runs unchanged on a laptop, in the cloud, and on leadership-class HPC systems. Because E4S publishes curated, versioned releases, a container tag plus a Spack specification is enough to reconstruct the exact software environment behind a published result — the same image and cache that produced a two-minute LAMMPS run on a laptop is the one that scales to a full simulation campaign.
CBIs are available for all major CPU and GPU platforms.
Getting Started
- Try it yourself: the E4S Quick Start walks through the same steps shown above — pull the image, supply code, launch, compile and run — with the E4S Guide Bot available for questions along the way.
- Browse available CBIs (CPU, CUDA, ROCm, OneAPI variants) on the Container Download page.
- Learn more about the build cache that backs every CBI: Use the E4S Spack Build Cache.
The Zero-to-LAMMPS benchmark above was developed as part of the PESO project (Partnering for Scientific-software Ecosystem Stewardship Opportunities).