[Clp] coinpackedmatrix appendCol taking a very long time

[David Prime]

Hi All,

I've been using CLP for a little while and am finding it to be very good,
proving fast and stable results.

One issue I've seen is that when I'm creating and executing many models in
parallel (we build and execute perhaps a few hundred models per second
across 16 cores in one program) the ratio of time spent creating the model
and then executing it goes from 1:1 to 10:1 under heavy load when we
attempt to run the same code on a 32 core machine with work spread evenly
across all cores. Execution times remain pretty much static, only
increasing with model complexity or when the machine running it is maxing
out all available cores and the program becomes CPU starved.

The model creation time is spent almost exclusively in the
CoinPackedMatrix->appendCol(
const int vecsize,
const int * vecind,
const double * vecelem )

method. I'm wondering if there's any internal global locks in any of the
libraries involved here that might be making performance worse as
parallelism increases. Another theory I have is that we only want to
increase parallelism from 16 to 32 cores when we're under very heavy load
so the characteristics of the models may also be changing in terms of the
size of the matrix we want to construct. Is there a more efficient way to
allocate the matrix than using appendCol?

Any thoughts?

Thanks,
David Prime
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://list.coin-or.org/pipermail/clp/attachments/20160926/63ba85bb/attachment.html>