Alpaca4d

Appearance

Solve large models on every core ​

An OpenSees built with Mumps factorises the stiffness matrix on every core of the machine. On a large model that makes an analysis 2 to 18 times quicker, and it needs less memory. On a small one it makes no difference.

A block of bricks in Von Mises stress, the solver speed test

The solver speed test: a block of bricks, solved with Mumps on every core.

You need ​

Steps ​

  1. In the Grasshopper menu bar, pick Alpaca4d → Settings → Set OpenSees Executable... and select the OpenSees built with Mumps. The path is kept for the next sessions.
  2. Leave System on Analysis Settings at its default, Auto. It solves with Mumps when the OpenSees has it, and with BandSPD on one core when it does not. Nothing else in the definition has to change.
  3. To solve on fewer cores than all of them, right-click Analysis Settings and pick a number under Cores for Mumps. All is the default. The choice is saved with the component.
  4. Run the analysis as before.

Check ​

Read the log output of Run Analysis. When the OpenSees in use has no Mumps, Auto writes this line in it, and the analysis solved on one core:

Alpaca4d: this OpenSees has no Mumps solver, so BandSPD solves, on one core

Without that line, Mumps solved.

Good to know ​

  • Set System to Mumps rather than Auto to be warned: on an OpenSees without Mumps the log then carries a warning, and BandSPD solves.
  • Cores for Mumps applies to Mumps, and to Auto on an OpenSees that has Mumps. Every other System solves on one core.
  • Natural Vibration needs no Analysis Settings. With its default solver, -genBandArpack, it uses Mumps on every core when the OpenSees has it.
  • Most of the gain comes from Mumps itself, not from the extra cores. Going from one core to all of them takes off up to a further quarter.

See also ​