Reduce Linux wheel size below 25 MB

[shimwell]

Summary

The Linux manylinux wheel built on the making-wheel-3 branch (PR #70) is 31.4 MB compressed (104.8 MB unpacked). This documents where the size goes and three concrete, low-risk levers to bring it under 25 MB.

Measured from openmc-0.15.3-cp312-cp312-manylinux_2_28_x86_64.whl (run 28036782175, commit e591c09).

Where the size goes (top contributors, compressed)

File	Compressed	Unpacked	Stripped?
libembree4.so.4	16.13 MB	46.60 MB	❌ not stripped
libMOAB.so.5.5.0	3.09 MB	16.10 MB	❌ not stripped
libhdf5.so	2.80 MB	10.39 MB	❌ not stripped
libnetcdf.so.22	1.52 MB	4.69 MB	❌ not stripped
libopenmc.so	1.35 MB	3.66 MB	✅ stripped
curl/TLS/DAP chain (21 libs)	4.20 MB	—	✅ mostly stripped

libopenmc.so and most direct deps are already stripped; the large bundled dependencies (Embree, MOAB, HDF5, netCDF) are not.

Three levers (need to cut > 6.4 MB)

1. Build Embree with fewer ISA variants — saves ~7–8 MB

tools/ci/cibw_before_all_linux.sh builds Embree v4.3.3 with all ISA kernels compiled in (SSE2, SSE4.2, AVX, AVX2, AVX-512). That is the bulk of the 16 MB — it is machine code, not debug info, so stripping does not help it.

Building only the ISAs we actually need roughly halves Embree. Suggested cmake flags (keeps the AVX-512 fast path and an AVX2 fallback for non-AVX-512 CPUs):

-DEMBREE_ISA_SSE2=OFF
-DEMBREE_ISA_SSE42=OFF   # (optionally keep =ON for pre-2013/no-AVX2 CPUs)
-DEMBREE_ISA_AVX=OFF
-DEMBREE_ISA_AVX2=ON
-DEMBREE_ISA_AVX512=ON

Tradeoff: AVX2+AVX-512 only drops support for pre-2013 (no-AVX2) CPUs in the ray tracer. Keeping EMBREE_ISA_SSE42=ON (3 ISAs) retains broad CPU support and still saves a large chunk.

2. Build netCDF without DAP — saves 4.2 MB (removes 21 libs)

tools/ci/cibw_before_all_linux.sh builds netCDF with -DENABLE_DAP=ON, which pulls in libcurl and its entire transitive TLS stack (libcrypto, libssl, libssh, libkrb5, libldap, libsasl2, libnghttp2, libidn2, libpsl, brotli, …) — 21 libraries totaling 4.20 MB compressed that auditwheel then bundles. OpenMC reads local mesh/HDF5 files only and never uses netCDF's remote (HTTP/DAP) access.

-DENABLE_DAP=OFF

3. Strip bundled libraries — saves ~2.2 MB

The auditwheel repair command in pyproject.toml does not strip:

repair-wheel-command = "auditwheel repair -w {dest_dir} {wheel}"

Add --strip so the un-stripped Embree/MOAB/HDF5/netCDF symbol tables are removed:

repair-wheel-command = "auditwheel repair --strip -w {dest_dir} {wheel}"

Expected result

Applied	Approx. wheel size
Today	31.4 MB
#3 strip only	~29 MB
#2 + #3	~25 MB (on the edge)
#1 + #2 + #3	~18 MB ✅

Levers #2 and #3 are zero-tradeoff. Lever #1 is the decisive one and aligns with the AVX-512 portability goal of the branch (AVX2 = the "without AVX-512" path, AVX-512 = the fast path).