rdelaage
/
postgresql
mirror of https://git.postgresql.org/git/postgresql.git
1
0
Fork 0
Code Issues Packages Projects Releases Wiki Activity

Get rid of artificial restriction on hash table sizes on Windows. 

The point of introducing the hash_mem_multiplier GUC was to let users
reproduce the old behavior of hash aggregation, i.e. that it could use
more than work_mem at need.  However, the implementation failed to get
the job done on Win64, where work_mem is clamped to 2GB to protect
various places that calculate memory sizes using "long int".  As
written, the same clamp was applied to hash_mem.  This resulted in
severe performance regressions for queries requiring a bit more than
2GB for hash aggregation, as they now spill to disk and there's no
way to stop that.

Getting rid of the work_mem restriction seems like a good idea, but
it's a big job and could not conceivably be back-patched.  However,
there's only a fairly small number of places that are concerned with
the hash_mem value, and it turns out to be possible to remove the
restriction there without too much code churn or any ABI breaks.
So, let's do that for now to fix the regression, and leave the
larger task for another day.

This patch does introduce a bit more infrastructure that should help
with the larger task, namely pg_bitutils.h support for working with
size_t values.

Per gripe from Laurent Hasson.  Back-patch to v13 where the
behavior change came in.

Discussion: https://postgr.es/m/997817.1627074924@sss.pgh.pa.us
Discussion: https://postgr.es/m/MN2PR15MB25601E80A9B6D1BA6F592B1985E39@MN2PR15MB2560.namprd15.prod.outlook.com
This commit is contained in:
Tom Lane 2021-07-25 14:02:27 -04:00
parent d9d8aa9bb9
commit 28d936031a
12 changed files with 153 additions and 111 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

6
src/backend/executor/execGrouping.c
View File

@ -165,14 +165,16 @@ BuildTupleHashTableExt(PlanState *parent,
{
TupleHashTable hashtable;
Size entrysize = sizeof(TupleHashEntryData) + additionalsize;
int hash_mem = get_hash_mem();
Size hash_mem_limit;
MemoryContext oldcontext;
bool allow_jit;
Assert(nbuckets > 0);
/* Limit initial table size request to not more than hash_mem */
nbuckets = Min(nbuckets, (long) ((hash_mem * 1024L) / entrysize));
hash_mem_limit = get_hash_memory_limit() / entrysize;
if (nbuckets > hash_mem_limit)
nbuckets = hash_mem_limit;
oldcontext = MemoryContextSwitchTo(metacxt);

42
src/backend/executor/nodeAgg.c
View File

@ -1802,15 +1802,15 @@ hash_agg_set_limits(double hashentrysize, double input_groups, int used_bits,
{
int npartitions;
Size partition_mem;
int hash_mem = get_hash_mem();
Size hash_mem_limit = get_hash_memory_limit();
/* if not expected to spill, use all of hash_mem */
if (input_groups * hashentrysize < hash_mem * 1024L)
if (input_groups * hashentrysize <= hash_mem_limit)
{
if (num_partitions != NULL)
*num_partitions = 0;
*mem_limit = hash_mem * 1024L;
*ngroups_limit = *mem_limit / hashentrysize;
*mem_limit = hash_mem_limit;
*ngroups_limit = hash_mem_limit / hashentrysize;
return;
}
@ -1835,10 +1835,10 @@ hash_agg_set_limits(double hashentrysize, double input_groups, int used_bits,
* minimum number of partitions, so we aren't going to dramatically exceed
* work mem anyway.
*/
if (hash_mem * 1024L > 4 * partition_mem)
*mem_limit = hash_mem * 1024L - partition_mem;
if (hash_mem_limit > 4 * partition_mem)
*mem_limit = hash_mem_limit - partition_mem;
else
*mem_limit = hash_mem * 1024L * 0.75;
*mem_limit = hash_mem_limit * 0.75;
if (*mem_limit > hashentrysize)
*ngroups_limit = *mem_limit / hashentrysize;
@ -1992,32 +1992,36 @@ static int
hash_choose_num_partitions(double input_groups, double hashentrysize,
int used_bits, int *log2_npartitions)
{
Size mem_wanted;
int partition_limit;
Size hash_mem_limit = get_hash_memory_limit();
double partition_limit;
double mem_wanted;
double dpartitions;
int npartitions;
int partition_bits;
int hash_mem = get_hash_mem();
/*
* Avoid creating so many partitions that the memory requirements of the
* open partition files are greater than 1/4 of hash_mem.
*/
partition_limit =
(hash_mem * 1024L * 0.25 - HASHAGG_READ_BUFFER_SIZE) /
(hash_mem_limit * 0.25 - HASHAGG_READ_BUFFER_SIZE) /
HASHAGG_WRITE_BUFFER_SIZE;
mem_wanted = HASHAGG_PARTITION_FACTOR * input_groups * hashentrysize;
/* make enough partitions so that each one is likely to fit in memory */
npartitions = 1 + (mem_wanted / (hash_mem * 1024L));
dpartitions = 1 + (mem_wanted / hash_mem_limit);
if (npartitions > partition_limit)
npartitions = partition_limit;
if (dpartitions > partition_limit)
dpartitions = partition_limit;
if (npartitions < HASHAGG_MIN_PARTITIONS)
npartitions = HASHAGG_MIN_PARTITIONS;
if (npartitions > HASHAGG_MAX_PARTITIONS)
npartitions = HASHAGG_MAX_PARTITIONS;
if (dpartitions < HASHAGG_MIN_PARTITIONS)
dpartitions = HASHAGG_MIN_PARTITIONS;
if (dpartitions > HASHAGG_MAX_PARTITIONS)
dpartitions = HASHAGG_MAX_PARTITIONS;
/* HASHAGG_MAX_PARTITIONS limit makes this safe */
npartitions = (int) dpartitions;
/* ceil(log2(npartitions)) */
partition_bits = my_log2(npartitions);
@ -2030,7 +2034,7 @@ hash_choose_num_partitions(double input_groups, double hashentrysize,
*log2_npartitions = partition_bits;
/* number of partitions will be a power of two */
npartitions = 1L << partition_bits;
npartitions = 1 << partition_bits;
return npartitions;
}

126
src/backend/executor/nodeHash.c
View File

@ -675,15 +675,12 @@ ExecChooseHashTableSize(double ntuples, int tupwidth, bool useskew,
{
int tupsize;
double inner_rel_bytes;
long bucket_bytes;
long hash_table_bytes;
long skew_table_bytes;
long max_pointers;
long mppow2;
size_t hash_table_bytes;
size_t bucket_bytes;
size_t max_pointers;
int nbatch = 1;
int nbuckets;
double dbuckets;
int hash_mem = get_hash_mem();
/* Force a plausible relation size if no info */
if (ntuples <= 0.0)
@ -700,9 +697,9 @@ ExecChooseHashTableSize(double ntuples, int tupwidth, bool useskew,
inner_rel_bytes = ntuples * tupsize;
/*
* Target in-memory hashtable size is hash_mem kilobytes.
* Compute in-memory hashtable size limit from GUCs.
*/
hash_table_bytes = hash_mem * 1024L;
hash_table_bytes = get_hash_memory_limit();
/*
* Parallel Hash tries to use the combined hash_mem of all workers to
@ -710,7 +707,14 @@ ExecChooseHashTableSize(double ntuples, int tupwidth, bool useskew,
* per worker and tries to process batches in parallel.
*/
if (try_combined_hash_mem)
hash_table_bytes += hash_table_bytes * parallel_workers;
{
/* Careful, this could overflow size_t */
double newlimit;
newlimit = (double) hash_table_bytes * (double) (parallel_workers + 1);
newlimit = Min(newlimit, (double) SIZE_MAX);
hash_table_bytes = (size_t) newlimit;
}
*space_allowed = hash_table_bytes;
@ -730,9 +734,12 @@ ExecChooseHashTableSize(double ntuples, int tupwidth, bool useskew,
*/
if (useskew)
{
skew_table_bytes = hash_table_bytes * SKEW_HASH_MEM_PERCENT / 100;
size_t bytes_per_mcv;
size_t skew_mcvs;
/*----------
* Compute number of MCVs we could hold in hash_table_bytes
*
* Divisor is:
* size of a hash tuple +
* worst-case size of skewBucket[] per MCV +
@ -740,12 +747,26 @@ ExecChooseHashTableSize(double ntuples, int tupwidth, bool useskew,
* size of skew bucket struct itself
*----------
*/
*num_skew_mcvs = skew_table_bytes / (tupsize +
(8 * sizeof(HashSkewBucket *)) +
sizeof(int) +
SKEW_BUCKET_OVERHEAD);
if (*num_skew_mcvs > 0)
hash_table_bytes -= skew_table_bytes;
bytes_per_mcv = tupsize +
(8 * sizeof(HashSkewBucket *)) +
sizeof(int) +
SKEW_BUCKET_OVERHEAD;
skew_mcvs = hash_table_bytes / bytes_per_mcv;
/*
* Now scale by SKEW_HASH_MEM_PERCENT (we do it in this order so as
* not to worry about size_t overflow in the multiplication)
*/
skew_mcvs = (skew_mcvs * SKEW_HASH_MEM_PERCENT) / 100;
/* Now clamp to integer range */
skew_mcvs = Min(skew_mcvs, INT_MAX);
*num_skew_mcvs = (int) skew_mcvs;
/* Reduce hash_table_bytes by the amount needed for the skew table */
if (skew_mcvs > 0)
hash_table_bytes -= skew_mcvs * bytes_per_mcv;
}
else
*num_skew_mcvs = 0;
@ -753,22 +774,20 @@ ExecChooseHashTableSize(double ntuples, int tupwidth, bool useskew,
/*
* Set nbuckets to achieve an average bucket load of NTUP_PER_BUCKET when
* memory is filled, assuming a single batch; but limit the value so that
* the pointer arrays we'll try to allocate do not exceed hash_mem nor
* MaxAllocSize.
* the pointer arrays we'll try to allocate do not exceed hash_table_bytes
* nor MaxAllocSize.
*
* Note that both nbuckets and nbatch must be powers of 2 to make
* ExecHashGetBucketAndBatch fast.
*/
max_pointers = *space_allowed / sizeof(HashJoinTuple);
max_pointers = hash_table_bytes / sizeof(HashJoinTuple);
max_pointers = Min(max_pointers, MaxAllocSize / sizeof(HashJoinTuple));
/* If max_pointers isn't a power of 2, must round it down to one */
mppow2 = 1L << my_log2(max_pointers);
if (max_pointers != mppow2)
max_pointers = mppow2 / 2;
max_pointers = pg_prevpower2_size_t(max_pointers);
/* Also ensure we avoid integer overflow in nbatch and nbuckets */
/* (this step is redundant given the current value of MaxAllocSize) */
max_pointers = Min(max_pointers, INT_MAX / 2);
max_pointers = Min(max_pointers, INT_MAX / 2 + 1);
dbuckets = ceil(ntuples / NTUP_PER_BUCKET);
dbuckets = Min(dbuckets, max_pointers);
@ -776,7 +795,7 @@ ExecChooseHashTableSize(double ntuples, int tupwidth, bool useskew,
/* don't let nbuckets be really small, though ... */
nbuckets = Max(nbuckets, 1024);
/* ... and force it to be a power of 2. */
nbuckets = 1 << my_log2(nbuckets);
nbuckets = pg_nextpower2_32(nbuckets);
/*
* If there's not enough space to store the projected number of tuples and
@ -786,10 +805,10 @@ ExecChooseHashTableSize(double ntuples, int tupwidth, bool useskew,
if (inner_rel_bytes + bucket_bytes > hash_table_bytes)
{
/* We'll need multiple batches */
long lbuckets;
size_t sbuckets;
double dbatch;
int minbatch;
long bucket_size;
size_t bucket_size;
/*
* If Parallel Hash with combined hash_mem would still need multiple
@ -813,10 +832,10 @@ ExecChooseHashTableSize(double ntuples, int tupwidth, bool useskew,
* overhead for the hash code, pointer to the next tuple, etc.
*/
bucket_size = (tupsize * NTUP_PER_BUCKET + sizeof(HashJoinTuple));
lbuckets = 1L << my_log2(hash_table_bytes / bucket_size);
lbuckets = Min(lbuckets, max_pointers);
nbuckets = (int) lbuckets;
nbuckets = 1 << my_log2(nbuckets);
sbuckets = pg_nextpower2_size_t(hash_table_bytes / bucket_size);
sbuckets = Min(sbuckets, max_pointers);
nbuckets = (int) sbuckets;
nbuckets = pg_nextpower2_32(nbuckets);
bucket_bytes = nbuckets * sizeof(HashJoinTuple);
/*
@ -1097,14 +1116,12 @@ ExecParallelHashIncreaseNumBatches(HashJoinTable hashtable)
/* Figure out how many batches to use. */
if (hashtable->nbatch == 1)
{
int hash_mem = get_hash_mem();
/*
* We are going from single-batch to multi-batch. We need
* to switch from one large combined memory budget to the
* regular hash_mem budget.
*/
pstate->space_allowed = hash_mem * 1024L;
pstate->space_allowed = get_hash_memory_limit();
/*
* The combined hash_mem of all participants wasn't
@ -1113,7 +1130,7 @@ ExecParallelHashIncreaseNumBatches(HashJoinTable hashtable)
* insufficient. So try two batches per participant,
* rounded up to a power of two.
*/
new_nbatch = 1 << my_log2(pstate->nparticipants * 2);
new_nbatch = pg_nextpower2_32(pstate->nparticipants * 2);
}
else
{
@ -1152,7 +1169,7 @@ ExecParallelHashIncreaseNumBatches(HashJoinTable hashtable)
MaxAllocSize / sizeof(dsa_pointer_atomic));
new_nbuckets = (int) dbuckets;
new_nbuckets = Max(new_nbuckets, 1024);
new_nbuckets = 1 << my_log2(new_nbuckets);
new_nbuckets = pg_nextpower2_32(new_nbuckets);
dsa_free(hashtable->area, old_batch0->buckets);
hashtable->batches[0].shared->buckets =
dsa_allocate(hashtable->area,
@ -3372,39 +3389,24 @@ ExecParallelHashTuplePrealloc(HashJoinTable hashtable, int batchno, size_t size)
}
/*
* Get a hash_mem value by multiplying the work_mem GUC's value by the
* hash_mem_multiplier GUC's value.
* Calculate the limit on how much memory can be used by Hash and similar
* plan types. This is work_mem times hash_mem_multiplier, and is
* expressed in bytes.
*
* Returns a work_mem style KB value that hash-based nodes (including but not
* limited to hash join) use in place of work_mem. This is subject to the
* same restrictions as work_mem itself. (There is no such thing as the
* hash_mem GUC, but it's convenient for our callers to pretend that there
* is.)
*
* Exported for use by the planner, as well as other hash-based executor
* Exported for use by the planner, as well as other hash-like executor
* nodes. This is a rather random place for this, but there is no better
* place.
*/
int
get_hash_mem(void)
size_t
get_hash_memory_limit(void)
{
double hash_mem;
double mem_limit;
Assert(hash_mem_multiplier >= 1.0);
/* Do initial calculation in double arithmetic */
mem_limit = (double) work_mem * hash_mem_multiplier * 1024.0;
hash_mem = (double) work_mem * hash_mem_multiplier;
/* Clamp in case it doesn't fit in size_t */
mem_limit = Min(mem_limit, (double) SIZE_MAX);
/*
* guc.c enforces a MAX_KILOBYTES limitation on work_mem in order to
* support the assumption that raw derived byte values can be stored in
* 'long' variables. The returned hash_mem value must also meet this
* assumption.
*
* We clamp the final value rather than throw an error because it should
* be possible to set work_mem and hash_mem_multiplier independently.
*/
if (hash_mem < MAX_KILOBYTES)
return (int) hash_mem;
return MAX_KILOBYTES;
return (size_t) mem_limit;
}

2
src/backend/executor/nodeMemoize.c
View File

@ -905,7 +905,7 @@ ExecInitMemoize(Memoize *node, EState *estate, int eflags)
mstate->mem_used = 0;
/* Limit the total memory consumed by the cache to this */
mstate->mem_limit = get_hash_mem() * 1024L;
mstate->mem_limit = get_hash_memory_limit();
/* A memory context dedicated for the cache */
mstate->tableContext = AllocSetContextCreate(CurrentMemoryContext,

7
src/backend/optimizer/path/costsize.c
View File

@ -2438,7 +2438,7 @@ cost_memoize_rescan(PlannerInfo *root, MemoizePath *mpath,
Cost total_cost;
/* available cache space */
hash_mem_bytes = get_hash_mem() * 1024L;
hash_mem_bytes = get_hash_memory_limit();
/*
* Set the number of bytes each cache entry should consume in the cache.
@ -3860,7 +3860,6 @@ final_cost_hashjoin(PlannerInfo *root, HashPath *path,
Cost run_cost = workspace->run_cost;
int numbuckets = workspace->numbuckets;
int numbatches = workspace->numbatches;
int hash_mem;
Cost cpu_per_tuple;
QualCost hash_qual_cost;
QualCost qp_qual_cost;
@ -3986,10 +3985,8 @@ final_cost_hashjoin(PlannerInfo *root, HashPath *path,
* that way, so it will be unable to drive the batch size below hash_mem
* when this is true.)
*/
hash_mem = get_hash_mem();
if (relation_byte_size(clamp_row_est(inner_path_rows * innermcvfreq),
inner_path->pathtarget->width) >
(hash_mem * 1024L))
inner_path->pathtarget->width) > get_hash_memory_limit())
startup_cost += disable_cost;
/*

6
src/backend/optimizer/plan/planner.c
View File

@ -3668,7 +3668,7 @@ consider_groupingsets_paths(PlannerInfo *root,
double dNumGroups)
{
Query *parse = root->parse;
int hash_mem = get_hash_mem();
Size hash_mem_limit = get_hash_memory_limit();
/*
* If we're not being offered sorted input, then only consider plans that
@ -3734,7 +3734,7 @@ consider_groupingsets_paths(PlannerInfo *root,
* with. Override hash_mem in that case; otherwise, we'll rely on the
* sorted-input case to generate usable mixed paths.
*/
if (hashsize > hash_mem * 1024L && gd->rollups)
if (hashsize > hash_mem_limit && gd->rollups)
return; /* nope, won't fit */
/*
@ -3853,7 +3853,7 @@ consider_groupingsets_paths(PlannerInfo *root,
{
List *rollups = NIL;
List *hash_sets = list_copy(gd->unsortable_sets);
double availspace = (hash_mem * 1024.0);
double availspace = hash_mem_limit;
ListCell *lc;
/*

6
src/backend/optimizer/plan/subselect.c
View File

@ -724,7 +724,6 @@ static bool
subplan_is_hashable(Plan *plan)
{
double subquery_size;
int hash_mem = get_hash_mem();
/*
* The estimated size of the subquery result must fit in hash_mem. (Note:
@ -734,7 +733,7 @@ subplan_is_hashable(Plan *plan)
*/
subquery_size = plan->plan_rows *
(MAXALIGN(plan->plan_width) + MAXALIGN(SizeofHeapTupleHeader));
if (subquery_size > hash_mem * 1024L)
if (subquery_size > get_hash_memory_limit())
return false;
return true;
@ -749,7 +748,6 @@ static bool
subpath_is_hashable(Path *path)
{
double subquery_size;
int hash_mem = get_hash_mem();
/*
* The estimated size of the subquery result must fit in hash_mem. (Note:
@ -759,7 +757,7 @@ subpath_is_hashable(Path *path)
*/
subquery_size = path->rows *
(MAXALIGN(path->pathtarget->width) + MAXALIGN(SizeofHeapTupleHeader));
if (subquery_size > hash_mem * 1024L)
if (subquery_size > get_hash_memory_limit())
return false;
return true;

8
src/backend/optimizer/prep/prepunion.c
View File

@ -1019,7 +1019,7 @@ choose_hashed_setop(PlannerInfo *root, List *groupClauses,
const char *construct)
{
int numGroupCols = list_length(groupClauses);
int hash_mem = get_hash_mem();
Size hash_mem_limit = get_hash_memory_limit();
bool can_sort;
bool can_hash;
Size hashentrysize;
@ -1055,13 +1055,11 @@ choose_hashed_setop(PlannerInfo *root, List *groupClauses,
*/
hashentrysize = MAXALIGN(input_path->pathtarget->width) + MAXALIGN(SizeofMinimalTupleHeader);
if (hashentrysize * dNumGroups > hash_mem * 1024L)
if (hashentrysize * dNumGroups > hash_mem_limit)
return false;
/*
* See if the estimated cost is no more than doing it the other way. We
* deliberately give the hash case more memory when hash_mem exceeds
* standard work mem (i.e. when hash_mem_multiplier exceeds 1.0).
* See if the estimated cost is no more than doing it the other way.
*
* We need to consider input_plan + hashagg versus input_plan + sort +
* group. Note that the actual result plan might involve a SetOp or

3
src/backend/optimizer/util/pathnode.c
View File

@ -1794,9 +1794,8 @@ create_unique_path(PlannerInfo *root, RelOptInfo *rel, Path *subpath,
* planner.c).
*/
int hashentrysize = subpath->pathtarget->width + 64;
int hash_mem = get_hash_mem();
if (hashentrysize * pathnode->path.rows > hash_mem * 1024L)
if (hashentrysize * pathnode->path.rows > get_hash_memory_limit())
{
/*
* We should not try to hash. Hack the SpecialJoinInfo to

6
src/backend/storage/ipc/shm_mq.c
View File

@ -727,11 +727,7 @@ shm_mq_receive(shm_mq_handle *mqh, Size *nbytesp, void **datap, bool nowait)
* Increase size to the next power of 2 that's >= nbytes, but
* limit to MaxAllocSize.
*/
#if SIZEOF_SIZE_T == 4
newbuflen = pg_nextpower2_32(nbytes);
#else
newbuflen = pg_nextpower2_64(nbytes);
#endif
newbuflen = pg_nextpower2_size_t(nbytes);
newbuflen = Min(newbuflen, MaxAllocSize);
if (mqh->mqh_buffer != NULL)

2
src/include/miscadmin.h
View File

@ -486,6 +486,6 @@ extern bool BackupInProgress(void);
extern void CancelBackup(void);
/* in executor/nodeHash.c */
extern int get_hash_mem(void);
extern size_t get_hash_memory_limit(void);
#endif /* MISCADMIN_H */

50
src/include/port/pg_bitutils.h
View File

@ -137,7 +137,7 @@ pg_rightmost_one_pos64(uint64 word)
/*
* pg_nextpower2_32
* Returns the next highest power of 2 of 'num', or 'num', if it's
* Returns the next higher power of 2 above 'num', or 'num' if it's
* already a power of 2.
*
* 'num' mustn't be 0 or be above PG_UINT32_MAX / 2 + 1.
@ -160,7 +160,7 @@ pg_nextpower2_32(uint32 num)
/*
* pg_nextpower2_64
* Returns the next highest power of 2 of 'num', or 'num', if it's
* Returns the next higher power of 2 above 'num', or 'num' if it's
* already a power of 2.
*
* 'num' mustn't be 0 or be above PG_UINT64_MAX / 2 + 1.
@ -181,6 +181,52 @@ pg_nextpower2_64(uint64 num)
return ((uint64) 1) << (pg_leftmost_one_pos64(num) + 1);
}
/*
* pg_nextpower2_size_t
* Returns the next higher power of 2 above 'num', for a size_t input.
*/
#if SIZEOF_SIZE_T == 4
#define pg_nextpower2_size_t(num) pg_nextpower2_32(num)
#else
#define pg_nextpower2_size_t(num) pg_nextpower2_64(num)
#endif
/*
* pg_prevpower2_32
* Returns the next lower power of 2 below 'num', or 'num' if it's
* already a power of 2.
*
* 'num' mustn't be 0.
*/
static inline uint32
pg_prevpower2_32(uint32 num)
{
return ((uint32) 1) << pg_leftmost_one_pos32(num);
}
/*
* pg_prevpower2_64
* Returns the next lower power of 2 below 'num', or 'num' if it's
* already a power of 2.
*
* 'num' mustn't be 0.
*/
static inline uint64
pg_prevpower2_64(uint64 num)
{
return ((uint64) 1) << pg_leftmost_one_pos64(num);
}
/*
* pg_prevpower2_size_t
* Returns the next lower power of 2 below 'num', for a size_t input.
*/
#if SIZEOF_SIZE_T == 4
#define pg_prevpower2_size_t(num) pg_prevpower2_32(num)
#else
#define pg_prevpower2_size_t(num) pg_prevpower2_64(num)
#endif
/*
* pg_ceil_log2_32
* Returns equivalent of ceil(log2(num))