\begin{flushleft}
{\bf 8. Conclusions}
\end{flushleft}

The  URM model has  been  successfully  parallelized and  applied to a
variety of computational environments. A significant  effort  has been
spent  on creating a uniform compilation and  execution  interface for
the model.  This was necessary to make the application portable across
various architectures.  The present organization allows  execution  to
be  reconfigured  from  a  large  supercomputer  to  a  set  of  local
workstations, and back,  with  one or  two commands.  This flexibility
makes  porting   the   application  to  new  architectures   easy  and
straightforward.  The procedures explored here are directly applicable
to other photochemical models as well.

Performance of  the application on various platforms differs depending
on network connections and memory  sharing  across those systems.  For
example, for application to Los Angeles a speed-up of 5.5 was achieved
using  the   Alpha  cluster   AXP's  which  are  connected  using  the
DEC$^\copyright$  Gigaswitch, whereas,  a speed  up of  only  3.8  was
achieved  using  CMU  cluster  AXP's which are  connected by a  slower
Ethernet network. Both cases used 10 processors in the simulations.  A
better speed-up (6.3  using 10 processors, and 3.7 using 5 processors)
was  achieved for application  to the  Northeast domain  because of the
larger grain size of the problem.  Performance measurements do suggest
that  there  is scope  for  substantial improvement  in  the future by
utilizing various  types of parallelism  that have  been  identified.
This  points to  the  future efforts of making the model
more scalable, and porting the  environmental models to new, massively
parallel architectures  (e.g.   Cray  T3D),  and  workstation  clusters
connected by even faster network [{\it Steenkiste et al.}, 1992].  The
computational power of  these new architectures will enable even  more
detailed simulations  which will  significantly increase  the level of
detail and accuracy in air quality modeling.


\begin{flushleft}
{\it Acknowledgements}
\end{flushleft}

This research was supported in part by the National Science Foundation
via  grant  ASC-9217365,  the  Environmental Protection  Agency  under
agreements  R819356  and CR-820908, and the Defense  Advanced Research
Projects  Agency  (DARPA)  under   Contract  MDA972-90-C-0035.   Their
support is gratefully appreciated.  We would  also like to acknowledge
the  use  of  computational  resources  provided   by  the  Pittsburgh
Supercomputing Center (PSC). We  would also  like to thank B. E.  Lara
for her valuable comments.