st: Mata performance monitoring

[Gordon Hughes <G.A.Hughes@ed.ac.uk>]

I am experiencing severe performance bottlenecks in implementing a 
maximum likelihood estimator for spatial panel data using Mata.  This 
is not surprising given the size of dataset (~ 200 panel units and 20 
time periods) and the complexity of the likelihood function which has 
to be implemented using numerical derivatives.

However, I would like to find out where the pay-off would be greatest 
for improving the code and reducing memory usage.  Some systems offer 
tools that permit the user to track how much time and memory is 
consumed for each call to functions and sub-programs, so that one can 
compile statistics of what operations represent potential 
bottlenecks.  In any ML program it is always calls to the likelihood 
function that use the most time, but I would like to find out whether 
significant time and/or memory is being used in other sub-programs.

I am not aware of similar tools for Mata or of ways in which the 
information could be compiled.  Can anyone suggest how I might do 
this?  As a corollary, is there any conventional wisdom about which 
of the optimisation methods is best under such 
circumstances.  Conventional wisdom in the past would have pointed to 
Newton-Raphson (nr) but this assumed that one could program 
analytical gradients and Hessian.  What is the most efficient for 
large datasets with numerical rather than analytical derivatives?

Gordon Hughes
g.a.hughes@ed.ac.uk

*
*   For searches and help try:
*   http://www.stata.com/help.cgi?search
*   http://www.stata.com/support/statalist/faq
*   http://www.ats.ucla.edu/stat/stata/