This milestone ties is with "Startup & Shutdown". [[br[[

In some instances, performance would be better if

  • a factorization is cached and later re-used by pivoting a few variables

instead of having the code

  • refactor from scratch

Enable the capability, explore the conditions under which it gives better performance than refactoring, and set the default parameters accordingly. <expand, pls!>

This milestone ties in with "startup & shutdown" (not to be confused with, "shut up & sit down.")

What information should remain in CLP and what (if any) information should be moved into OSI? <pls elaborate, give examples, etc.>

Identify the matrix structure. Choose a dual or primal approach depending on the ratio of rows to columns.

<details?> <if this has already been done, explain what was done and throw out any ideas that are ripe for exploration.>

Improve the performance of CLP for network problems.

<details?> <if this is already done, please explain what's been done & throw out ideas that could be explored>

Decide whether or not you can re-order the rows and/or columns to improve performance. Rumor has it that Anshul Gupta did some work long this line. <elaborate>

Improving performance as the code "starts up & shuts up" is a relatively self-contained piece of work that needs to be done. There are several different components to it.

  1. arrays <more details here!>
  2. scaling <more details here!>
  3. factorization, including speeding up the factorization for problems with less than 10K rows <more details here!>
  4. put back parallel capability (of factorization?) and do it well ("multi-core machines are on the way" - LL)<more details here!>

On thought is that JJHF could/would/should lead a discussion, share his thoughts, get feedback, and be the lead on the effort, but be assisted by others who had interest in getting more familiar with the source code. One of the design issues to be addressed is what information needs to be interfaced through OSI.

Add cability to recognize and exploit "subnetwork" structure in the problem.

Note: See TracRoadmap for help on using the roadmap.