Here is a clean version of the patch. Step 1 is an optimization. Step 2
refactors this:
HTAB *
ShmemInitHash(const char *name, /* table string name for shmem index */
- long init_size, /* initial table size */
+ long init_size, /* initial table size XXX ignored,
refactor! */
As I understand there is no performance degradation:
Before
======
Core i7, pgbench -j 8 -c 8 -T 50 pgbench:
```
number of transactions actually processed: 14315
latency average: 27.945 ms
latency stddev: 44.920 ms
tps = 286.097043 (including connections establishing)
tps = 286.127642 (excluding connections establishing)
```
60-core server, pgbench -j 64 -c 64 -T 50 pgbench:
```
number of transactions actually processed: 176127
latency average: 18.162 ms
latency stddev: 23.861 ms
tps = 3521.069672 (including connections establishing)
tps = 3522.166706 (excluding connections establishing)
```
After
=====
Core i7, pgbench -j 8 -c 8 -T 50 pgbench:
```
number of transactions actually processed: 14212
latency average: 27.984 ms
latency stddev: 43.250 ms
tps = 284.038588 (including connections establishing)
tps = 285.112772 (excluding connections establishing)
```
60-core server, pgbench -j 64 -c 64 -T 50 pgbench:
```
number of transactions actually processed: 172135
latency average: 17.767 ms
latency stddev: 22.692 ms
tps = 3590.410302 (including connections establishing)
tps = 3594.607165 (excluding connections establishing)
```
diff --git a/src/backend/storage/ipc/shmem.c b/src/backend/storage/ipc/shmem.c
index 78f15f0..d361dbb 100644
--- a/src/backend/storage/ipc/shmem.c
+++ b/src/backend/storage/ipc/shmem.c
@@ -265,7 +265,7 @@ InitShmemIndex(void)
*/
HTAB *
ShmemInitHash(const char *name, /* table string name for shmem index */
- long init_size, /* initial table size */
+ long init_size, /* initial table size XXX ignored, refactor! */
long max_size, /* max size of the table */
HASHCTL *infoP, /* info about key and bucket size */
int hash_flags) /* info about infoP */
@@ -299,7 +299,7 @@ ShmemInitHash(const char *name, /* table string name for shmem index */
/* Pass location of hashtable header to hash_create */
infoP->hctl = (HASHHDR *) location;
- return hash_create(name, init_size, infoP, hash_flags);
+ return hash_create(name, max_size, infoP, hash_flags);
}
/*
diff --git a/src/backend/utils/hash/dynahash.c b/src/backend/utils/hash/dynahash.c
index eacffc4..a50a2d3 100644
--- a/src/backend/utils/hash/dynahash.c
+++ b/src/backend/utils/hash/dynahash.c
@@ -15,7 +15,7 @@
* to hash_create. This prevents any attempt to split buckets on-the-fly.
* Therefore, each hash bucket chain operates independently, and no fields
* of the hash header change after init except nentries and freeList.
- * A partitioned table uses a spinlock to guard changes of those two fields.
+ * A partitioned table uses spinlocks to guard changes of those fields.
* This lets any subset of the hash buckets be treated as a separately
* lockable partition. We expect callers to use the low-order bits of a
* lookup key's hash value as a partition number --- this will work because
@@ -87,6 +87,7 @@
#include "access/xact.h"
#include "storage/shmem.h"
#include "storage/spin.h"
+#include "storage/lock.h"
#include "utils/dynahash.h"
#include "utils/memutils.h"
@@ -128,12 +129,21 @@ typedef HASHBUCKET *HASHSEGMENT;
*/
struct HASHHDR
{
- /* In a partitioned table, take this lock to touch nentries or freeList */
- slock_t mutex; /* unused if not partitioned table */
+ /*
+ * In a partitioned table take these locks to touch nentries or freeList.
+ * If hash table is not partitioned mutexes are not used.
+ */
+ slock_t mutex[NUM_LOCK_PARTITIONS];
+
+ /*
+ * These fields change during entry addition/deletion. If hash table is
+ * not partitioned only nentries[0] and freeList[0] are used.
+ */
- /* These fields change during entry addition/deletion */
- long nentries; /* number of entries in hash table */
- HASHELEMENT *freeList; /* linked list of free elements */
+ /* number of entries in hash table */
+ long nentries[NUM_LOCK_PARTITIONS];
+ /* linked list of free elements */
+ HASHELEMENT *freeList[NUM_LOCK_PARTITIONS];
/* These fields can change, but not in a partitioned table */
/* Also, dsize can't change in a shared table, even if unpartitioned */
@@ -166,6 +176,8 @@ struct HASHHDR
#define IS_PARTITIONED(hctl) ((hctl)->num_partitions != 0)
+#define PARTITION_IDX(hctl, hashcode) (IS_PARTITIONED(hctl) ? LockHashPartition(hashcode) : 0)
+
/*
* Top control structure for a hashtable --- in a shared table, each backend
* has its own copy (OK since no fields change at runtime)
@@ -219,10 +231,10 @@ static long hash_accesses,
*/
static void *DynaHashAlloc(Size size);
static HASHSEGMENT seg_alloc(HTAB *hashp);
-static bool element_alloc(HTAB *hashp, int nelem);
+static bool element_alloc(HTAB *hashp, int nelem, int partition_idx);
static bool dir_realloc(HTAB *hashp);
static bool expand_table(HTAB *hashp);
-static HASHBUCKET get_hash_entry(HTAB *hashp);
+static HASHBUCKET get_hash_entry(HTAB *hashp, int partition_idx);
static void hdefault(HTAB *hashp);
static int choose_nelem_alloc(Size entrysize);
static bool init_htab(HTAB *hashp, long nelem);
@@ -282,6 +294,9 @@ hash_create(const char *tabname, long nelem, HASHCTL *info, int flags)
{
HTAB *hashp;
HASHHDR *hctl;
+ int i,
+ partitions_number,
+ nelem_alloc;
/*
* For shared hash tables, we have a local hash header (HTAB struct) that
@@ -482,10 +497,20 @@ hash_create(const char *tabname, long nelem, HASHCTL *info, int flags)
if ((flags & HASH_SHARED_MEM) ||
nelem < hctl->nelem_alloc)
{
- if (!element_alloc(hashp, (int) nelem))
- ereport(ERROR,
- (errcode(ERRCODE_OUT_OF_MEMORY),
- errmsg("out of memory")));
+ if (IS_PARTITIONED(hashp->hctl))
+ partitions_number = NUM_LOCK_PARTITIONS;
+ else
+ partitions_number = 1;
+
+ nelem_alloc = ((int) nelem) / partitions_number;
+ if (nelem_alloc == 0)
+ nelem_alloc = 1;
+
+ for (i = 0; i < partitions_number; i++)
+ if (!element_alloc(hashp, nelem_alloc, i))
+ ereport(ERROR,
+ (errcode(ERRCODE_OUT_OF_MEMORY),
+ errmsg("out of memory")));
}
if (flags & HASH_FIXED_SIZE)
@@ -503,9 +528,6 @@ hdefault(HTAB *hashp)
MemSet(hctl, 0, sizeof(HASHHDR));
- hctl->nentries = 0;
- hctl->freeList = NULL;
-
hctl->dsize = DEF_DIRSIZE;
hctl->nsegs = 0;
@@ -572,12 +594,14 @@ init_htab(HTAB *hashp, long nelem)
HASHSEGMENT *segp;
int nbuckets;
int nsegs;
+ int i;
/*
* initialize mutex if it's a partitioned table
*/
if (IS_PARTITIONED(hctl))
- SpinLockInit(&hctl->mutex);
+ for (i = 0; i < NUM_LOCK_PARTITIONS; i++)
+ SpinLockInit(&(hctl->mutex[i]));
/*
* Divide number of elements by the fill factor to determine a desired
@@ -648,7 +672,7 @@ init_htab(HTAB *hashp, long nelem)
"HIGH MASK ", hctl->high_mask,
"LOW MASK ", hctl->low_mask,
"NSEGS ", hctl->nsegs,
- "NENTRIES ", hctl->nentries);
+ "NENTRIES ", hash_get_num_entries(hctl));
#endif
return true;
}
@@ -769,7 +793,7 @@ hash_stats(const char *where, HTAB *hashp)
where, hashp->hctl->accesses, hashp->hctl->collisions);
fprintf(stderr, "hash_stats: entries %ld keysize %ld maxp %u segmentcount %ld\n",
- hashp->hctl->nentries, (long) hashp->hctl->keysize,
+ hash_get_num_entries(hashp), (long) hashp->hctl->keysize,
hashp->hctl->max_bucket, hashp->hctl->nsegs);
fprintf(stderr, "%s: total accesses %ld total collisions %ld\n",
where, hash_accesses, hash_collisions);
@@ -863,6 +887,7 @@ hash_search_with_hash_value(HTAB *hashp,
HASHBUCKET currBucket;
HASHBUCKET *prevBucketPtr;
HashCompareFunc match;
+ int partition_idx = PARTITION_IDX(hctl, hashvalue);
#if HASH_STATISTICS
hash_accesses++;
@@ -885,7 +910,7 @@ hash_search_with_hash_value(HTAB *hashp,
* order of these tests is to try to check cheaper conditions first.
*/
if (!IS_PARTITIONED(hctl) && !hashp->frozen &&
- hctl->nentries / (long) (hctl->max_bucket + 1) >= hctl->ffactor &&
+ hctl->nentries[0] / (long) (hctl->max_bucket + 1) >= hctl->ffactor &&
!has_seq_scans(hashp))
(void) expand_table(hashp);
}
@@ -943,20 +968,20 @@ hash_search_with_hash_value(HTAB *hashp,
{
/* if partitioned, must lock to touch nentries and freeList */
if (IS_PARTITIONED(hctl))
- SpinLockAcquire(&hctl->mutex);
+ SpinLockAcquire(&(hctl->mutex[partition_idx]));
- Assert(hctl->nentries > 0);
- hctl->nentries--;
+ Assert(hctl->nentries[partition_idx] > 0);
+ hctl->nentries[partition_idx]--;
/* remove record from hash bucket's chain. */
*prevBucketPtr = currBucket->link;
/* add the record to the freelist for this table. */
- currBucket->link = hctl->freeList;
- hctl->freeList = currBucket;
+ currBucket->link = hctl->freeList[partition_idx];
+ hctl->freeList[partition_idx] = currBucket;
if (IS_PARTITIONED(hctl))
- SpinLockRelease(&hctl->mutex);
+ SpinLockRelease(&hctl->mutex[partition_idx]);
/*
* better hope the caller is synchronizing access to this
@@ -982,7 +1007,7 @@ hash_search_with_hash_value(HTAB *hashp,
elog(ERROR, "cannot insert into frozen hashtable \"%s\"",
hashp->tabname);
- currBucket = get_hash_entry(hashp);
+ currBucket = get_hash_entry(hashp, partition_idx);
if (currBucket == NULL)
{
/* out of memory */
@@ -1175,41 +1200,71 @@ hash_update_hash_key(HTAB *hashp,
* create a new entry if possible
*/
static HASHBUCKET
-get_hash_entry(HTAB *hashp)
+get_hash_entry(HTAB *hashp, int partition_idx)
{
- HASHHDR *hctl = hashp->hctl;
+ HASHHDR *hctl = hashp->hctl;
HASHBUCKET newElement;
+ int i,
+ borrow_from_idx;
for (;;)
{
/* if partitioned, must lock to touch nentries and freeList */
if (IS_PARTITIONED(hctl))
- SpinLockAcquire(&hctl->mutex);
+ SpinLockAcquire(&hctl->mutex[partition_idx]);
/* try to get an entry from the freelist */
- newElement = hctl->freeList;
+ newElement = hctl->freeList[partition_idx];
+
if (newElement != NULL)
- break;
+ {
+ /* remove entry from freelist, bump nentries */
+ hctl->freeList[partition_idx] = newElement->link;
+ hctl->nentries[partition_idx]++;
+ if (IS_PARTITIONED(hctl))
+ SpinLockRelease(&hctl->mutex[partition_idx]);
+
+ return newElement;
+ }
- /* no free elements. allocate another chunk of buckets */
if (IS_PARTITIONED(hctl))
- SpinLockRelease(&hctl->mutex);
+ SpinLockRelease(&hctl->mutex[partition_idx]);
- if (!element_alloc(hashp, hctl->nelem_alloc))
+ /* no free elements. allocate another chunk of buckets */
+ if (!element_alloc(hashp, hctl->nelem_alloc, partition_idx))
{
- /* out of memory */
- return NULL;
- }
- }
+ if (!IS_PARTITIONED(hctl))
+ return NULL; /* out of memory */
- /* remove entry from freelist, bump nentries */
- hctl->freeList = newElement->link;
- hctl->nentries++;
+ /* try to borrow element from another partition */
+ borrow_from_idx = partition_idx;
+ for (;;)
+ {
+ borrow_from_idx = (borrow_from_idx + 1) % NUM_LOCK_PARTITIONS;
+ if (borrow_from_idx == partition_idx)
+ break;
- if (IS_PARTITIONED(hctl))
- SpinLockRelease(&hctl->mutex);
+ SpinLockAcquire(&(hctl->mutex[borrow_from_idx]));
+ newElement = hctl->freeList[borrow_from_idx];
+
+ if (newElement != NULL)
+ {
+ hctl->freeList[borrow_from_idx] = newElement->link;
+ SpinLockRelease(&(hctl->mutex[borrow_from_idx]));
+
+ SpinLockAcquire(&hctl->mutex[partition_idx]);
+ hctl->nentries[partition_idx]++;
+ SpinLockRelease(&hctl->mutex[partition_idx]);
+
+ break;
+ }
- return newElement;
+ SpinLockRelease(&(hctl->mutex[borrow_from_idx]));
+ }
+
+ return newElement;
+ }
+ }
}
/*
@@ -1218,11 +1273,21 @@ get_hash_entry(HTAB *hashp)
long
hash_get_num_entries(HTAB *hashp)
{
+ int i;
+ long sum = hashp->hctl->nentries[0];
+
/*
* We currently don't bother with the mutex; it's only sensible to call
* this function if you've got lock on all partitions of the table.
*/
- return hashp->hctl->nentries;
+
+ if (!IS_PARTITIONED(hashp->hctl))
+ return sum;
+
+ for (i = 1; i < NUM_LOCK_PARTITIONS; i++)
+ sum += hashp->hctl->nentries[i];
+
+ return sum;
}
/*
@@ -1530,9 +1595,9 @@ seg_alloc(HTAB *hashp)
* allocate some new elements and link them into the free list
*/
static bool
-element_alloc(HTAB *hashp, int nelem)
+element_alloc(HTAB *hashp, int nelem, int partition_idx)
{
- HASHHDR *hctl = hashp->hctl;
+ HASHHDR *hctl = hashp->hctl;
Size elementSize;
HASHELEMENT *firstElement;
HASHELEMENT *tmpElement;
@@ -1563,14 +1628,14 @@ element_alloc(HTAB *hashp, int nelem)
/* if partitioned, must lock to touch freeList */
if (IS_PARTITIONED(hctl))
- SpinLockAcquire(&hctl->mutex);
+ SpinLockAcquire(&hctl->mutex[partition_idx]);
/* freelist could be nonempty if two backends did this concurrently */
- firstElement->link = hctl->freeList;
- hctl->freeList = prevElement;
+ firstElement->link = hctl->freeList[partition_idx];
+ hctl->freeList[partition_idx] = prevElement;
if (IS_PARTITIONED(hctl))
- SpinLockRelease(&hctl->mutex);
+ SpinLockRelease(&hctl->mutex[partition_idx]);
return true;
}
diff --git a/src/include/storage/lwlock.h b/src/include/storage/lwlock.h
index ff34529..6e866f1 100644
--- a/src/include/storage/lwlock.h
+++ b/src/include/storage/lwlock.h
@@ -128,13 +128,13 @@ extern char *MainLWLockNames[];
* having this file include lock.h or bufmgr.h would be backwards.
*/
-/* Number of partitions of the shared buffer mapping hashtable */
-#define NUM_BUFFER_PARTITIONS 128
-
/* Number of partitions the shared lock tables are divided into */
-#define LOG2_NUM_LOCK_PARTITIONS 4
+#define LOG2_NUM_LOCK_PARTITIONS 7
#define NUM_LOCK_PARTITIONS (1 << LOG2_NUM_LOCK_PARTITIONS)
+/* Number of partitions of the shared buffer mapping hashtable */
+#define NUM_BUFFER_PARTITIONS NUM_LOCK_PARTITIONS
+
/* Number of partitions the shared predicate lock tables are divided into */
#define LOG2_NUM_PREDICATELOCK_PARTITIONS 4
#define NUM_PREDICATELOCK_PARTITIONS (1 << LOG2_NUM_PREDICATELOCK_PARTITIONS)
diff --git a/contrib/pg_stat_statements/pg_stat_statements.c b/contrib/pg_stat_statements/pg_stat_statements.c
index cea84ee..656afc3 100644
--- a/contrib/pg_stat_statements/pg_stat_statements.c
+++ b/contrib/pg_stat_statements/pg_stat_statements.c
@@ -495,7 +495,7 @@ pgss_shmem_startup(void)
info.hash = pgss_hash_fn;
info.match = pgss_match_fn;
pgss_hash = ShmemInitHash("pg_stat_statements hash",
- pgss_max, pgss_max,
+ pgss_max,
&info,
HASH_ELEM | HASH_FUNCTION | HASH_COMPARE);
diff --git a/src/backend/storage/buffer/buf_table.c b/src/backend/storage/buffer/buf_table.c
index 6ed47d5..3d5aec7 100644
--- a/src/backend/storage/buffer/buf_table.c
+++ b/src/backend/storage/buffer/buf_table.c
@@ -62,7 +62,7 @@ InitBufTable(int size)
info.num_partitions = NUM_BUFFER_PARTITIONS;
SharedBufHash = ShmemInitHash("Shared Buffer Lookup Table",
- size, size,
+ size,
&info,
HASH_ELEM | HASH_BLOBS | HASH_PARTITION);
}
diff --git a/src/backend/storage/ipc/shmem.c b/src/backend/storage/ipc/shmem.c
index d361dbb..7c254ec 100644
--- a/src/backend/storage/ipc/shmem.c
+++ b/src/backend/storage/ipc/shmem.c
@@ -237,7 +237,7 @@ InitShmemIndex(void)
hash_flags = HASH_ELEM;
ShmemIndex = ShmemInitHash("ShmemIndex",
- SHMEM_INDEX_SIZE, SHMEM_INDEX_SIZE,
+ SHMEM_INDEX_SIZE,
&info, hash_flags);
}
@@ -255,17 +255,12 @@ InitShmemIndex(void)
* exceeded substantially (since it's used to compute directory size and
* the hash table buckets will get overfull).
*
- * init_size is the number of hashtable entries to preallocate. For a table
- * whose maximum size is certain, this should be equal to max_size; that
- * ensures that no run-time out-of-shared-memory failures can occur.
- *
* Note: before Postgres 9.0, this function returned NULL for some failure
* cases. Now, it always throws error instead, so callers need not check
* for NULL.
*/
HTAB *
ShmemInitHash(const char *name, /* table string name for shmem index */
- long init_size, /* initial table size XXX ignored, refactor! */
long max_size, /* max size of the table */
HASHCTL *infoP, /* info about key and bucket size */
int hash_flags) /* info about infoP */
diff --git a/src/backend/storage/lmgr/lock.c b/src/backend/storage/lmgr/lock.c
index 76fc615..378b54a 100644
--- a/src/backend/storage/lmgr/lock.c
+++ b/src/backend/storage/lmgr/lock.c
@@ -373,8 +373,7 @@ void
InitLocks(void)
{
HASHCTL info;
- long init_table_size,
- max_table_size;
+ long max_table_size;
bool found;
/*
@@ -382,7 +381,6 @@ InitLocks(void)
* calculations must agree with LockShmemSize!
*/
max_table_size = NLOCKENTS();
- init_table_size = max_table_size / 2;
/*
* Allocate hash table for LOCK structs. This stores per-locked-object
@@ -394,14 +392,12 @@ InitLocks(void)
info.num_partitions = NUM_LOCK_PARTITIONS;
LockMethodLockHash = ShmemInitHash("LOCK hash",
- init_table_size,
max_table_size,
&info,
HASH_ELEM | HASH_BLOBS | HASH_PARTITION);
/* Assume an average of 2 holders per lock */
max_table_size *= 2;
- init_table_size *= 2;
/*
* Allocate hash table for PROCLOCK structs. This stores
@@ -413,7 +409,6 @@ InitLocks(void)
info.num_partitions = NUM_LOCK_PARTITIONS;
LockMethodProcLockHash = ShmemInitHash("PROCLOCK hash",
- init_table_size,
max_table_size,
&info,
HASH_ELEM | HASH_FUNCTION | HASH_PARTITION);
diff --git a/src/backend/storage/lmgr/predicate.c b/src/backend/storage/lmgr/predicate.c
index 4b4ad81..7898d81 100644
--- a/src/backend/storage/lmgr/predicate.c
+++ b/src/backend/storage/lmgr/predicate.c
@@ -1116,7 +1116,6 @@ InitPredicateLocks(void)
PredicateLockTargetHash = ShmemInitHash("PREDICATELOCKTARGET hash",
max_table_size,
- max_table_size,
&info,
HASH_ELEM | HASH_BLOBS |
HASH_PARTITION | HASH_FIXED_SIZE);
@@ -1144,7 +1143,6 @@ InitPredicateLocks(void)
PredicateLockHash = ShmemInitHash("PREDICATELOCK hash",
max_table_size,
- max_table_size,
&info,
HASH_ELEM | HASH_FUNCTION |
HASH_PARTITION | HASH_FIXED_SIZE);
@@ -1225,7 +1223,6 @@ InitPredicateLocks(void)
SerializableXidHash = ShmemInitHash("SERIALIZABLEXID hash",
max_table_size,
- max_table_size,
&info,
HASH_ELEM | HASH_BLOBS |
HASH_FIXED_SIZE);
diff --git a/src/include/storage/shmem.h b/src/include/storage/shmem.h
index c94d620..a361302 100644
--- a/src/include/storage/shmem.h
+++ b/src/include/storage/shmem.h
@@ -37,7 +37,7 @@ extern void InitShmemAllocation(void);
extern void *ShmemAlloc(Size size);
extern bool ShmemAddrIsValid(const void *addr);
extern void InitShmemIndex(void);
-extern HTAB *ShmemInitHash(const char *name, long init_size, long max_size,
+extern HTAB *ShmemInitHash(const char *name, long max_size,
HASHCTL *infoP, int hash_flags);
extern void *ShmemInitStruct(const char *name, Size size, bool *foundPtr);
extern Size add_size(Size s1, Size s2);
--
Sent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)
To make changes to your subscription:
http://www.postgresql.org/mailpref/pgsql-hackers
