From 4631acb275d09de01f8b297ad4d7708ddeed9e8f Mon Sep 17 00:00:00 2001 From: amitlan Date: Tue, 15 Jun 2021 16:21:48 +0900 Subject: [PATCH v7] Teach get_partition_for_tuple to cache bound offset For bulk loads into list and range partitioned tables, it can be very likely that long runs of consecutive tuples route to the same partition. In such cases, we can avoid the overhead of performing a binary search for each such tuple by caching the offset of the bound for that partition and checking that the bound indeed satisfies any subsequent tuples, which can be implemented with fewer comparisons than the binary search. To avoid impacting the cases where such caching can be unproductive, an adaptive algorithm is used to determine whether to actually enable caching or to disable it if checking the cached offset seems to add pure overhead per tuple. Author: Hou Zhijie Author: Amit Langote --- src/backend/executor/execPartition.c | 214 ++++++++++++++++++++++++--- 1 file changed, 192 insertions(+), 22 deletions(-) diff --git a/src/backend/executor/execPartition.c b/src/backend/executor/execPartition.c index 606c920b06..d52f8c4931 100644 --- a/src/backend/executor/execPartition.c +++ b/src/backend/executor/execPartition.c @@ -133,6 +133,13 @@ struct PartitionTupleRouting * routing it through this table). A NULL value is stored if no tuple * conversion is required. * + * cached_bound_offset + * last_seen_offset + * n_tups_inserted + * n_offset_changed + * Fields to manage the state for bound offset caching; see + * maybe_cache_partition_bound_offset() + * * indexes * Array of partdesc->nparts elements. For leaf partitions the index * corresponds to the partition's ResultRelInfo in the encapsulating @@ -150,6 +157,12 @@ typedef struct PartitionDispatchData PartitionDesc partdesc; TupleTableSlot *tupslot; AttrMap *tupmap; + + int cached_bound_offset; + int last_seen_offset; + int n_tups_inserted; + int n_offset_changed; + int indexes[FLEXIBLE_ARRAY_MEMBER]; } PartitionDispatchData; @@ -1026,6 +1039,10 @@ ExecInitPartitionDispatchInfo(EState *estate, pd->key = RelationGetPartitionKey(rel); pd->keystate = NIL; pd->partdesc = partdesc; + + pd->cached_bound_offset = pd->last_seen_offset = -1; + pd->n_tups_inserted = pd->n_offset_changed = 0; + if (parent_pd != NULL) { TupleDesc tupdesc = RelationGetDescr(rel); @@ -1231,6 +1248,135 @@ FormPartitionKeyDatum(PartitionDispatch pd, elog(ERROR, "wrong number of partition key expressions"); } +/* + * Threshold of the number of tuples to need to have been processed before + * maybe_cache_partition_bound_offset() (re-)assesses whether caching must be + * enabled or disabled for subsequent tuples. + */ +#define CACHE_BOUND_OFFSET_THRESHOLD_TUPS 1000 + +/* + * maybe_cache_partition_bound_offset + * Conditionally sets pd->cached_bound_offset so that + * get_cached_{list|range}_partition can be used for subsequent + * tuples + * + * It is set if it appears that some offsets observed over the last + * pd->n_tups_inserted tuples would have been reused, which can be inferred + * from seeing that the ratio of tuples inserted and the number of times the + * offset needed to be changed during the insertion of those tuples is greater + * than 1. Conversely, we disable the caching if the ratio is 1, because that + * suggests that every consecutively inserted tuple mapped to a different + * partition. + */ +static inline void +maybe_cache_partition_bound_offset(PartitionDispatch pd, int offset) +{ + if (offset != pd->last_seen_offset) + { + pd->last_seen_offset = offset; + pd->n_offset_changed += 1; + /* Only set to the new value after calculating the ratio formula. */ + pd->cached_bound_offset = -1; + } + + /* + * Only consider (re-)enabling/disabling caching if we've seen at least + * a threshold number of tuples since the last time we enabled/disabled + * it. + */ + if (pd->n_tups_inserted < CACHE_BOUND_OFFSET_THRESHOLD_TUPS) + return; + + /* Wouldn't get called if the cached bound offset worked. */ + Assert(offset != pd->cached_bound_offset); + + /* If the offset didn't change at all, caching it might be a good idea. */ + if (pd->n_offset_changed == 0 || + (double) pd->n_tups_inserted / pd->n_offset_changed > 1) + pd->cached_bound_offset = offset; + else + pd->cached_bound_offset = -1; + pd->n_tups_inserted = pd->n_offset_changed = 0; +} + +/* + * get_cached_{list|range}_partition + * Computes if the cached bound offset value, if any, is satisfied by + * the tuple specified in 'values' and it is, returns the index of + * the partition corresponding to that bound + * + * Callers must ensure that none of the elements of 'values' is NULL. + */ +static inline int +get_cached_list_partition(PartitionDispatch pd, + PartitionBoundInfo boundinfo, + PartitionKey key, + Datum *values) +{ + int part_index = -1; + int cached_off = pd->cached_bound_offset; + + if (cached_off >= 0) + { + Datum bound_datum = boundinfo->datums[cached_off][0]; + int32 cmpval; + + cmpval = DatumGetInt32(FunctionCall2Coll(&key->partsupfunc[0], + key->partcollation[0], + bound_datum, + values[0])); + if (cmpval == 0) + part_index = boundinfo->indexes[cached_off]; + } + + return part_index; +} + +static inline int +get_cached_range_partition(PartitionDispatch pd, + PartitionBoundInfo boundinfo, + PartitionKey key, + Datum *values) +{ + int part_index = -1; + int cached_off = pd->cached_bound_offset; + + if (cached_off >= 0) + { + Datum *bound_datums = boundinfo->datums[cached_off]; + PartitionRangeDatumKind *bound_kind = boundinfo->kind[cached_off]; + int32 cmpval; + + /* Check if the value is above the low bound */ + cmpval = partition_rbound_datum_cmp(key->partsupfunc, + key->partcollation, + bound_datums, + bound_kind, + values, + key->partnatts); + if (cmpval == 0) + part_index = boundinfo->indexes[cached_off + 1]; + else if (cmpval < 0 && cached_off + 1 < boundinfo->ndatums) + { + /* Check if the value is below the high bound */ + bound_datums = boundinfo->datums[cached_off + 1]; + bound_kind = boundinfo->kind[cached_off + 1]; + cmpval = partition_rbound_datum_cmp(key->partsupfunc, + key->partcollation, + bound_datums, + bound_kind, + values, + key->partnatts); + + if (cmpval > 0) + part_index = boundinfo->indexes[cached_off + 1]; + } + } + + return part_index; +} + /* * get_partition_for_tuple * Finds partition of relation which accepts the partition key specified @@ -1248,6 +1394,8 @@ get_partition_for_tuple(PartitionDispatch pd, Datum *values, bool *isnull) PartitionDesc partdesc = pd->partdesc; PartitionBoundInfo boundinfo = partdesc->boundinfo; + pd->n_tups_inserted += 1; + /* Route as appropriate based on partitioning strategy. */ switch (key->strategy) { @@ -1272,14 +1420,26 @@ get_partition_for_tuple(PartitionDispatch pd, Datum *values, bool *isnull) } else { - bool equal = false; - - bound_offset = partition_list_bsearch(key->partsupfunc, - key->partcollation, - boundinfo, - values[0], &equal); - if (bound_offset >= 0 && equal) - part_index = boundinfo->indexes[bound_offset]; + part_index = get_cached_list_partition(pd, + boundinfo, + key, + values); + if (part_index < 0) + { + bool equal = false; + + bound_offset = partition_list_bsearch(key->partsupfunc, + key->partcollation, + boundinfo, + values[0], &equal); + if (bound_offset >= 0 && equal) + { + part_index = boundinfo->indexes[bound_offset]; + if (part_index >= 0) + maybe_cache_partition_bound_offset(pd, + bound_offset); + } + } } break; @@ -1304,20 +1464,30 @@ get_partition_for_tuple(PartitionDispatch pd, Datum *values, bool *isnull) if (!range_partkey_has_null) { - bound_offset = partition_range_datum_bsearch(key->partsupfunc, - key->partcollation, - boundinfo, - key->partnatts, - values, - &equal); - - /* - * The bound at bound_offset is less than or equal to the - * tuple value, so the bound at offset+1 is the upper - * bound of the partition we're looking for, if there - * actually exists one. - */ - part_index = boundinfo->indexes[bound_offset + 1]; + part_index = get_cached_range_partition(pd, + boundinfo, + key, + values); + if (part_index < 0) + { + bound_offset = partition_range_datum_bsearch(key->partsupfunc, + key->partcollation, + boundinfo, + key->partnatts, + values, + &equal); + + /* + * The bound at bound_offset is less than or equal to the + * tuple value, so the bound at offset+1 is the upper + * bound of the partition we're looking for, if there + * actually exists one. + */ + part_index = boundinfo->indexes[bound_offset + 1]; + if (part_index >= 0) + maybe_cache_partition_bound_offset(pd, + bound_offset); + } } } break; -- 2.24.1