Changeset: da7578b49268 for MonetDB
URL: http://dev.monetdb.org/hg/MonetDB?cmd=changeset;node=da7578b49268
Modified Files:
        MonetDB5/src/optimizer/opt_tarantula.mx
Branch: default
Log Message:

Just textual


diffs (41 lines):

diff -r d4f39a8bca27 -r da7578b49268 MonetDB5/src/optimizer/opt_tarantula.mx
--- a/MonetDB5/src/optimizer/opt_tarantula.mx   Fri Nov 12 22:28:11 2010 +0100
+++ b/MonetDB5/src/optimizer/opt_tarantula.mx   Sat Nov 13 10:31:46 2010 +0100
@@ -170,6 +170,17 @@
 an upcall to that context (=dangerous).
 The solution is that any variable context should be
 passed through a relation.
+
+[The hard world] The way plans are generated by mitosis leads to a large 
number of
+mat.pack() operations, whose result is subsequently spread out over the legs 
again
+for continual processing. This leads to a lot of transport, from 
leg->head->all-legs.
+This situation can be circumvented by allowing each leg to call upon all 
tarantula legs
+to solve the problem at hand. With recycling enabled in each leg, duplicate 
work will
+be avoided and a direct copy of the data is obtained.
+
+To make this work, we have to keep track of what input variables for a leg
+are effectively the result of a tarantula call and inject its materialisation
+in the leg upon first use.
 @{
 @mal
 module tarantula; 
@@ -931,14 +942,17 @@
        int otop[MAXSLICES], output[MAXSLICES][MAXSHARE];
        InstrPtr *list;
        int *needed;
-       char *done, *msg= MAL_SUCCEED;
+       char *done;
 
-       /* merge table may leave some mat.new() dead code */
+       /* merge table may leave some mat.new() dead code 
+       should be handled by the pipes
+       char *msg= MAL_SUCCEED;
        msg = OPTdeadcode(cntxt, mb, stk, 0);
        if ( msg){
                GDKfree(msg);
                return 0;
        }
+       */
        if( cntxt == 0){
                /* confuscate, delay for later activation */
                TARinitcode(cntxt, mb);
_______________________________________________
Checkin-list mailing list
[email protected]
http://mail.monetdb.org/mailman/listinfo/checkin-list

Reply via email to