[kepler-dev] Using PN Directors in Kepler

[Dan Higgins]

Perhaps most Kepler users understand this, but I just now managed to 
understand some of the details of using PN directors and thought I would 
try to document some issues.

My inital problem came when I tried to replace the SDF director in the 
simple GARP demo workflow (kepler\workflows\eco\garpModel_ImageJ.xml). 
This workflow has 2 Constant actors, the 3 GARP actors (JNI), and an 
ImageJ output display actor. When I replaces the SDF director by a PN 
director, Kepler either crashed or continued to execute forever.

One method for getting this workflow to run under PN is shown in the 
slightly modified version 
(kepler\workflows\eco\garpModel_ImageJ_PN.xml). The required changes are 
the replacement of the SDF director with a PN director and the addition 
of a Ramp actor before the 2 Constant actors with the Ramp actor set to 
fire only once.

Why does this work? The real problem is the Constant actors used to 
supply string inputs (file names) to the GARP actors. The ordinary 
Constant actor will continually supply tokens to whatever actor it is 
connected to.

As I understand it, all actors in a workflow are in separate theads. 
Actors receive fire events (in some order I don't understand) but they 
'block' on the 'get()' method for getting the input port tokens. This 
means that the code in the 'fire' method of actors will not execute 
until 'get' actually has an input token on all ports where it is looking 
for tokens. And, of course, it will not put output tokens on output 
ports until the end of the fire method. PN execution ends when all ports 
are 'blocked'; i.e. there are no inputs.

This explains the problem with the GARP workflow. When the SDF director 
is just replaced with the PN director, the Constant actors supply tokens 
to the GARP PresampleLayers actor and it begins to run. But the PN 
director will continue to supply tokens, so that GARP actor keeps 
getting fired even before the first instance has completed. Adding the 
Ramp actor makes sure the Contant actors are only fired once, so the PN 
director doesn't keep trying to fire the GARP actor (and after all the 
actors are finished the execution terminates).

[Note that Efrat recognized this problem some time ago and created the 
ConstOnce actor. This actor just extends the Constant actor and returns 
false from the PostFire event to stop returning tokens.]

-- 
*******************************************************************
Dan Higgins                                  higgins at nceas.ucsb.edu
http://www.nceas.ucsb.edu/    Ph: 805-893-5127
National Center for Ecological Analysis and Synthesis (NCEAS) 
Marine Science Building - Room 3405
Santa Barbara, CA 93195
*******************************************************************