Changeset: ce23fa286bb6 for MonetDB
URL: https://dev.monetdb.org/hg/MonetDB?cmd=changeset;node=ce23fa286bb6
Modified Files:
gdk/gdk_join.c
Branch: Jun2020
Log Message:
Ignore the cost of creating a hash when the bat is persistent.
The idea is that the hash is going to be used a lot of times, so the
cost of creating it once doesn't matter too much.
diffs (67 lines):
diff --git a/gdk/gdk_join.c b/gdk/gdk_join.c
--- a/gdk/gdk_join.c
+++ b/gdk/gdk_join.c
@@ -3162,7 +3162,13 @@ leftjoin(BAT **r1p, BAT **r2p, BAT *l, B
/* no hash table, so cost includes time to build the
* hash table (single scan) plus the time to do the
* lookups (also single scan, we assume some chains) */
- rcost = (double) rci.ncand * 2 + lci.ncand * 1.1;
+ rcost = lci.ncand * 1.1;
+#ifdef PERSISTENTHASH
+ /* only count the cost of creating the hash for
+ * non-persistent bats */
+ if (rci.ncand != BATcount(r) || !(BBP_status(r->batCacheid) &
BBPEXISTING) || r->theap.dirty || GDKinmemory())
+#endif
+ rcost += rci.ncand * 2.0;
} else {
if (rci.noids > 0) {
/* if we need to do binary search on candidate
@@ -3205,7 +3211,13 @@ leftjoin(BAT **r1p, BAT **r2p, BAT *l, B
/* no hash table, so cost includes time to build the
* hash table (single scan) plus the time to do the
* lookups (also single scan, we assume some chains) */
- lcost = (double) lci.ncand * 2 + rci.ncand * 1.1;
+ lcost = rci.ncand * 1.1;
+#ifdef PERSISTENTHASH
+ /* only count the cost of creating the hash
+ * for non-persistent bats */
+ if (lci.ncand != BATcount(l) ||
!(BBP_status(l->batCacheid) & BBPEXISTING) || l->theap.dirty || GDKinmemory())
+#endif
+ lcost += lci.ncand * 2.0;
} else {
if (lci.noids > 0) {
/* if we need to do binary search on candidate
@@ -3501,9 +3513,13 @@ BATjoin(BAT **r1p, BAT **r2p, BAT *l, BA
/* no hash table, so cost includes time to build the
* hash table (single scan) plus the time to do the
* lookups (also single scan, we assume some chains) */
- lcost = (double) lci.ncand * 2 + rci.ncand * 1.1;
- if (lci.ncand == BATcount(l) && !l->batTransient)
- lcost *= 0.8;
+ lcost = rci.ncand * 1.1;
+#ifdef PERSISTENTHASH
+ /* only count the cost of creating the hash for
+ * non-persistent bats */
+ if (lci.ncand != BATcount(l) || !(BBP_status(l->batCacheid) &
BBPEXISTING) || l->theap.dirty || GDKinmemory())
+#endif
+ lcost += lci.ncand * 2.0;
} else {
if (lci.noids > 0) {
/* if we need to do binary search on candidate
@@ -3542,9 +3558,13 @@ BATjoin(BAT **r1p, BAT **r2p, BAT *l, BA
/* no hash table, so cost includes time to build the
* hash table (single scan) plus the time to do the
* lookups (also single scan, we assume some chains) */
- rcost = (double) rci.ncand * 2 + lci.ncand * 1.1;
- if (rci.ncand == BATcount(r) && !r->batTransient)
- rcost *= 0.8;
+ rcost = lci.ncand * 1.1;
+#ifdef PERSISTENTHASH
+ /* only count the cost of creating the hash for
+ * non-persistent bats */
+ if (rci.ncand != BATcount(r) || !(BBP_status(r->batCacheid) &
BBPEXISTING) || r->theap.dirty || GDKinmemory())
+#endif
+ rcost += rci.ncand * 2.0;
} else {
if (rci.noids > 0) {
/* if we need to do binary search on candidate
_______________________________________________
checkin-list mailing list
checkin-list@monetdb.org
https://www.monetdb.org/mailman/listinfo/checkin-list
