[Coin-discuss] New stable versions of Clp and Cbc and Parallel Cbc in beta test

[John J Forrest]

Since we moved COIN to svn there have been three versions of Clp and Cbc -
stable, trunk and branches/devel.  This has proved unwieldy and other
software such as Bonmin has depended on having branches/devel which is not
a satisfactory affair.  In fact having other people depend on
branches/devel has meant that I have kept my own private copy!

For Clp a new Clp/stable/1.4 has been created from the old Clp/trunk.
Clp/trunk now represents the old Clp/branches/devel.  The use of
Clp/branches/devel is now deprecated and no projects should use it.

For Cbc a new Cbc/stable/1.2 has been created from the old Cbc/trunk.
Cbc/trunk now represents the old public Cbc/branches/devel.  Other projects
should now point to Cbc/trunk.

Cbc/branches/devel (or a renamed version) will exist for a short time.
This now has my private copy which is enabled for parallel use of Cbc.
This is designed for use on SMP processors and I would like feedback over
the next month before moving to Cbc/trunk.  To try this version in
Cbc/branches/devel use --enable-cbc-parallel in the configure.  Parallel
use is asked for by setNumberThreads(n) or -thread n in cbc where n >1.
For normal use just node solution (with associated cuts and heuristics) go
parallel so for easy problems there is not much effect.  There are options
to do initial cuts at root in parallel and to go parallel in some
heuristics at root, but neither of these options have seemed to do much
good.  Here are results for "cbc -thread n -miplib" on my dual Cpu laptop
when runs were done three times for most n.

Threads     Elapsed        Cpu
      0       10:13      10:06
      2        6:24       9:20
      2        6:34       9:35
      2        7:45      11:45
      3        7:17      11:14
      3        8:43      13:49
      3        6:51      10:15
      4        7:09      10:56
      4        8:10      12:59
      4        7:27      11:38
     12!       7:45      13:48

These results are about what I would expect given the size of the problems.
Some of the problem times can vary a great deal with tiny changes and a
different order of node selection e.g. one problem time varied from 15
seconds to 150 seconds with 3 threads.

John Forrest