MATLAB's outofprocess mode has massive overhead

[ischoegl]

Problem description

While the outofprocess mode introduced in #1868 resolves #1852, it comes at the cost of massive overhead due to repeated allocation/copying of arrays, see #1997 (comment). The migration to the generated CLib in #1997 may resolve crashes reported in #1852 when using LD_PRELOAD.

Behavior

After #1868, the Cantera toolbox defaults to outofprocess regardless of operating system. The MATLAB samples ignite.m, ignite_uv.m, ignite_hp.m and plug_flow_reactor.m have extreme performance penalties as they rely on integration with MATLAB's ode15s solver, and only use Cantera for the evaluation of properties and kinetics. Recent testing in #1974 showed runtimes for plug_flow_reactor.m on Windows of around 30 minutes (!). PS: With some of the problematic examples disabled, Windows runners are significantly slower than macOS and Linux actions.

System information

• Cantera version: 3.2.0b1
• OS: any (but more pronounced on Windows)
• Python/MATLAB/other software versions: MATLAB 2025a

Additional context

• Utilize std::span & migrate to C++20 enhancements#237 will only help marginally with this issue, as CLib only copies arrays used for setup. Most CLib array operations are currently handled by raw pointers.

I opened this issue to document previously reported behavior. Tagging @ssun30 and @rwest.

[ssun30]

There's not much we can do about the inter-process communications latency, so I went for the low-hanging fruit and implemented caching where I can expect close to 100% hit rate. The most we can save is about 75% but we will still be an order of magnitude slower than inprocess mode.