[SOSlib-discuss] tcc-based online compilation, increasing SOSlib performance

[Rainer M. <ra...@tb...>]

Hi Everyone,

Great news:

as recently several people started to use SOSlib on supercomputing 
clusters, we were motivated to go into the code again and optimize.

Christoph and me went through Andrew Finney's old code for online 
compilation of the right-hand side functions passed to CVODES (including 
command-line calls etc.).

So I just managed to get the same code running under Linux using the 
Tiny C Compiler as a library and the very cool TCC online compilation:

   tcc_compile_string(TCCState *s, char *sourceCode);


So far it works only for the one bottle-neck function that evaluates 
ODEs (and not e.g. for jacobi evaluation or sensitivity stuff).

But that already brings significant performance gains! E.g. for the 
example model MAPK.xml, integrating to time 100000 in 1000 steps:

without compilation:  2.13 seconds
with tcc compilation: 0.53 seconds

A strange effect is that at higher error tolerances and long integration 
times the results start to differ. We need to check where the different 
round-offs etc. are made. At lower tolerances the results are exactly 
the same.


The new code is in CVS now. Linking to TCC is optional, so everything 
else should still work as before.

Caveat: TCC so far only works on 32-bit architecture, and thus all 
libraries (sundials, xerces/expat, libsbml and soslib) must also be 
compiled and run on 32-bit machines.

So we can only hope that TCC will soon be available for x86_64, 
otherwise we could still implement command-line calls, like for the 
Windows version (where back then tcc didn't work as a library).
(TCC can be run on 64-bit but only with -m32 options and asfaik it's not 
possible to have a mixed architecture code in on tool).


TCC is available from http://fabrice.bellard.free.fr/tcc/

However, I used Robert Landley's fork because I couldn't compile the 
original, and he seems to be the only one still working on it:

http://landley.net/code/tinycc

It should work with both ... in principle.


So if that all works (with all rhs functions), I think there are not 
much more possibilities to get faster, i.e. to make better use of 
SUNDIALS from arbitrary ODE models.

We can only include the usual reaction network analysis, i.e. mass 
conservation analysis, to get rid of unnecessary ODEs etc. However, I 
guess, that won't bring very much gain, because the lost equation is 
similar to a still existing one which still needs to be solved by 
CVODES. Maybe "fast reaction analysis" would have some potential?

So, I think we are close to becoming as fast as possible.

Well, there'd probably still be possibilities with model-specific 
"preconditioning" and such stuff about which I don't know much. Feel 
free to propose your ideas :)

Rainer