Changeset: 9d81c7b5a874 for MonetDB
URL: http://dev.monetdb.org/hg/MonetDB?cmd=changeset;node=9d81c7b5a874
Modified Files:
MonetDB5/src/mal/mal_instruction.mx
MonetDB5/src/mal/mal_interpreter.mx
MonetDB5/src/mal/mal_stack.mx
Branch: default
Log Message:
Hooks for admission control
The dataflow block execution has been extended with a hook for
admission and wrapup. They take effect when the instruction is
about to be added to or released from the queue.
The hooks context is a single stack frame.
These hooks are intended to finetune the scheduling using
more knowlegde about the context, or to implement a dynamic
dataflow crawler where the instructions to be executed are
determined by e.g. sampling first.
The stack structure is changed, so a clean rebuild may be required.
diffs (91 lines):
diff -r 3ce2ea819aea -r 9d81c7b5a874 MonetDB5/src/mal/mal_instruction.mx
--- a/MonetDB5/src/mal/mal_instruction.mx Wed Jun 30 21:27:45 2010 +0200
+++ b/MonetDB5/src/mal/mal_instruction.mx Thu Jul 01 09:09:21 2010 +0200
@@ -350,7 +350,6 @@
#define VARRETS 2
/* all functions return a string */
-typedef str (*MALfcn) ();
typedef struct {
bit token; /* instruction type */
diff -r 3ce2ea819aea -r 9d81c7b5a874 MonetDB5/src/mal/mal_interpreter.mx
--- a/MonetDB5/src/mal/mal_interpreter.mx Wed Jun 30 21:27:45 2010 +0200
+++ b/MonetDB5/src/mal/mal_interpreter.mx Thu Jul 01 09:09:21 2010 +0200
@@ -1262,7 +1262,15 @@
printInstruction(GDKstdout, flow->mb, 0, p, LIST_MAL_STMT);
}
}
-
+...@-
+Parallel processing is mostly driven by dataflow, but within this context
+there may be different schemes to take instructions into execution.
+The admission scheme (and wrapup) are the necessary scheduler hooks.
+A scheduler registers the functions needed and should release them
+at the end of the parallel block.
+They take effect after we have ensured that the basic properties for
+execution hold.
+...@c
static str
DFLOWscheduler( DataFlow flow )
{
@@ -1287,6 +1295,8 @@
while(queued){
PARDEBUG stream_printf(GDKstdout,"#waiting for results, queued
%d\n", queued);
f = q_dequeue(flow->done);
+ if ( f->flow->stk->wrapup ) /* clean up whatever is called for
*/
+ (*f->flow->stk->wrapup)(f->flow->cntxt, f->flow->mb,
f->flow->stk, getInstrPtr(flow->mb, f->pc));
f->status = DFLOWwrapup;
queued--;
if (f->pc < 0) {
@@ -1295,8 +1305,11 @@
}
/* we have to wait for all threads to report back */
/* dequeue the remainders in case of an error */
- while(queued-- > 0)
- (void)q_dequeue(flow->done);
+ while(queued-- > 0) {
+ if ( f->flow->stk->wrapup ) /* clean up
whatever is called for */
+
(*f->flow->stk->wrapup)(f->flow->cntxt, f->flow->mb, f->flow->stk,
getInstrPtr(flow->mb, f->pc));
+ (void)q_dequeue(flow->done);
+ }
return f->error;
} else {
p = getInstrPtr(flow->mb, f->pc);
@@ -1337,7 +1350,9 @@
if (fs[i].status == DFLOWpending ) {
p = getInstrPtr(flow->mb, fs[i].pc);
for ( j= p->retc; j < p->argc; j++)
- if ( getArg(p,j)== oa && DFLOWeligible(flow,fs,i,p,pc)){
+ if ( getArg(p,j)== oa && DFLOWeligible(flow,fs,i,p,pc)
&&
+ ( f->flow->stk->admit == 0 ||
(*f->flow->stk->admit)(f->flow->cntxt, f->flow->mb, f->flow->stk,
getInstrPtr(flow->mb, f->pc)) )
+ ){
queued++;
candidates ++;
DFLOWactivate(flow,fs,i,p);
diff -r 3ce2ea819aea -r 9d81c7b5a874 MonetDB5/src/mal/mal_stack.mx
--- a/MonetDB5/src/mal/mal_stack.mx Wed Jun 30 21:27:45 2010 +0200
+++ b/MonetDB5/src/mal/mal_stack.mx Thu Jul 01 09:09:21 2010 +0200
@@ -68,6 +68,8 @@
#define MAXGLOBALS 4 * STACKINCR
#define MAXSHARES 8
+typedef str (*MALfcn) ();
+
typedef struct MALSTK {
int stksize;
int stktop;
@@ -76,6 +78,13 @@
int calldepth; /* to protect against runtime stack overflow */
short keepAlive; /* do not garbage collect when set */
short garbageCollect; /* stack needs garbage collection */
+...@-
+Parallel processing is mostly driven by dataflow, but within this context
+there may be different schemes to take instructions into execution.
+The admission scheme (and wrapup) are the necessary scheduler hooks.
+...@h
+ MALfcn admit;
+ MALfcn wrapup;
MT_Lock stklock; /* used for parallel processing */
@-
It is handy to administer the timing in the stack frame
_______________________________________________
Checkin-list mailing list
Checkin-list@monetdb.org
http://mail.monetdb.org/mailman/listinfo/checkin-list
