agora inbox for [email protected]
help / color / mirror / Atom feedFrom: Jehan-Guillaume de Rorthais <[email protected]>
Subject: [PATCH] Allocate hash batches related BufFile in a dedicated context
Date: Mon, 27 Mar 2023 15:54:39 +0200
---
src/backend/executor/nodeHash.c | 43 +++++++++++++++++++++++++----
src/backend/executor/nodeHashjoin.c | 18 ++++++++----
src/include/executor/hashjoin.h | 15 ++++++++--
src/include/executor/nodeHashjoin.h | 2 +-
4 files changed, 64 insertions(+), 14 deletions(-)
diff --git a/src/backend/executor/nodeHash.c b/src/backend/executor/nodeHash.c
index 748c9b0024..3da83ac22a 100644
--- a/src/backend/executor/nodeHash.c
+++ b/src/backend/executor/nodeHash.c
@@ -484,7 +484,7 @@ ExecHashTableCreate(HashState *state, List *hashOperators, List *hashCollations,
*
* The hashtable control block is just palloc'd from the executor's
* per-query memory context. Everything else should be kept inside the
- * subsidiary hashCxt or batchCxt.
+ * subsidiary hashCxt, batchCxt or fileCxt.
*/
hashtable = palloc_object(HashJoinTableData);
hashtable->nbuckets = nbuckets;
@@ -538,6 +538,10 @@ ExecHashTableCreate(HashState *state, List *hashOperators, List *hashCollations,
"HashBatchContext",
ALLOCSET_DEFAULT_SIZES);
+ hashtable->fileCxt = AllocSetContextCreate(CurrentMemoryContext,
+ "HashBatchFiles",
+ ALLOCSET_DEFAULT_SIZES);
+
/* Allocate data that will live for the life of the hashjoin */
oldcxt = MemoryContextSwitchTo(hashtable->hashCxt);
@@ -570,15 +574,21 @@ ExecHashTableCreate(HashState *state, List *hashOperators, List *hashCollations,
if (nbatch > 1 && hashtable->parallel_state == NULL)
{
+ MemoryContext oldctx;
+
/*
* allocate and initialize the file arrays in hashCxt (not needed for
* parallel case which uses shared tuplestores instead of raw files)
*/
+ oldctx = MemoryContextSwitchTo(hashtable->fileCxt);
+
hashtable->innerBatchFile = palloc0_array(BufFile *, nbatch);
hashtable->outerBatchFile = palloc0_array(BufFile *, nbatch);
/* The files will not be opened until needed... */
/* ... but make sure we have temp tablespaces established for them */
PrepareTempTablespaces();
+
+ MemoryContextSwitchTo(oldctx);
}
MemoryContextSwitchTo(oldcxt);
@@ -929,12 +939,18 @@ ExecHashIncreaseNumBatches(HashJoinTable hashtable)
nbatch = oldnbatch * 2;
Assert(nbatch > 1);
+ elog(WARNING, "increasing number of batches from %d to %d", oldnbatch, nbatch);
+
+ elog(LOG, "ExecHashIncreaseNumBatches ======= context stats start =======");
+ MemoryContextStats(TopMemoryContext);
+
+
#ifdef HJDEBUG
printf("Hashjoin %p: increasing nbatch to %d because space = %zu\n",
hashtable, nbatch, hashtable->spaceUsed);
#endif
- oldcxt = MemoryContextSwitchTo(hashtable->hashCxt);
+ oldcxt = MemoryContextSwitchTo(hashtable->fileCxt);
if (hashtable->innerBatchFile == NULL)
{
@@ -1022,9 +1038,11 @@ ExecHashIncreaseNumBatches(HashJoinTable hashtable)
{
/* dump it out */
Assert(batchno > curbatch);
+
ExecHashJoinSaveTuple(HJTUPLE_MINTUPLE(hashTuple),
hashTuple->hashvalue,
- &hashtable->innerBatchFile[batchno]);
+ &hashtable->innerBatchFile[batchno],
+ hashtable->fileCxt);
hashtable->spaceUsed -= hashTupleSize;
nfreed++;
@@ -1042,6 +1060,13 @@ ExecHashIncreaseNumBatches(HashJoinTable hashtable)
oldchunks = nextchunk;
}
+ if (hashtable->fileCxt->mem_allocated > hashtable->spaceAllowed)
+ elog(NOTICE,
+ "Growing number of hash batch to %d is exhausting allowed memory (%ld > %ld)",
+ nbatch,
+ hashtable->fileCxt->mem_allocated,
+ hashtable->spaceAllowed);
+
#ifdef HJDEBUG
printf("Hashjoin %p: freed %ld of %ld tuples, space now %zu\n",
hashtable, nfreed, ninmemory, hashtable->spaceUsed);
@@ -1063,6 +1088,9 @@ ExecHashIncreaseNumBatches(HashJoinTable hashtable)
hashtable);
#endif
}
+
+ elog(LOG, "ExecHashIncreaseNumBatches ======= context stats end =======");
+ MemoryContextStats(TopMemoryContext);
}
/*
@@ -1681,9 +1709,11 @@ ExecHashTableInsert(HashJoinTable hashtable,
* put the tuple into a temp file for later batches
*/
Assert(batchno > hashtable->curbatch);
+
ExecHashJoinSaveTuple(tuple,
hashvalue,
- &hashtable->innerBatchFile[batchno]);
+ &hashtable->innerBatchFile[batchno],
+ hashtable->fileCxt);
}
if (shouldFree)
@@ -2534,8 +2564,11 @@ ExecHashRemoveNextSkewBucket(HashJoinTable hashtable)
{
/* Put the tuple into a temp file for later batches */
Assert(batchno > hashtable->curbatch);
+
ExecHashJoinSaveTuple(tuple, hashvalue,
- &hashtable->innerBatchFile[batchno]);
+ &hashtable->innerBatchFile[batchno],
+ hashtable->fileCxt);
+
pfree(hashTuple);
hashtable->spaceUsed -= tupleSize;
hashtable->spaceUsedSkew -= tupleSize;
diff --git a/src/backend/executor/nodeHashjoin.c b/src/backend/executor/nodeHashjoin.c
index f189fb4d28..6055abde49 100644
--- a/src/backend/executor/nodeHashjoin.c
+++ b/src/backend/executor/nodeHashjoin.c
@@ -432,8 +432,10 @@ ExecHashJoinImpl(PlanState *pstate, bool parallel)
*/
Assert(parallel_state == NULL);
Assert(batchno > hashtable->curbatch);
+
ExecHashJoinSaveTuple(mintuple, hashvalue,
- &hashtable->outerBatchFile[batchno]);
+ &hashtable->outerBatchFile[batchno],
+ hashtable->fileCxt);
if (shouldFree)
heap_free_minimal_tuple(mintuple);
@@ -1234,21 +1236,27 @@ ExecParallelHashJoinNewBatch(HashJoinState *hjstate)
* The data recorded in the file for each tuple is its hash value,
* then the tuple in MinimalTuple format.
*
- * Note: it is important always to call this in the regular executor
- * context, not in a shorter-lived context; else the temp file buffers
- * will get messed up.
+ * Note: it is important always to call this in the HashBatchFiles context,
+ * not in a shorter-lived context; else the temp file buffers will get messed
+ * up.
*/
void
ExecHashJoinSaveTuple(MinimalTuple tuple, uint32 hashvalue,
- BufFile **fileptr)
+ BufFile **fileptr, MemoryContext filecxt)
{
BufFile *file = *fileptr;
if (file == NULL)
{
+ MemoryContext oldctx;
+
+ oldctx = MemoryContextSwitchTo(filecxt);
+
/* First write to this batch file, so open it. */
file = BufFileCreateTemp(false);
*fileptr = file;
+
+ MemoryContextSwitchTo(oldctx);
}
BufFileWrite(file, &hashvalue, sizeof(uint32));
diff --git a/src/include/executor/hashjoin.h b/src/include/executor/hashjoin.h
index acb7592ca0..d759235d7f 100644
--- a/src/include/executor/hashjoin.h
+++ b/src/include/executor/hashjoin.h
@@ -25,10 +25,14 @@
*
* Each active hashjoin has a HashJoinTable control block, which is
* palloc'd in the executor's per-query context. All other storage needed
- * for the hashjoin is kept in private memory contexts, two for each hashjoin.
+ * for the hashjoin is kept in private memory contexts, three for each
+ * hashjoin:
+ * - HashTableContext (hashCxt): the control block associated to the hash table
+ * - HashBatchContext (batchCxt): storages for batches
+ * - HashBatchFiles (fileCxt): storage for temp files buffers
+ *
* This makes it easy and fast to release the storage when we don't need it
- * anymore. (Exception: data associated with the temp files lives in the
- * per-query context too, since we always call buffile.c in that context.)
+ * anymore.
*
* The hashtable contexts are made children of the per-query context, ensuring
* that they will be discarded at end of statement even if the join is
@@ -39,6 +43,10 @@
* "hashCxt", while storage that is only wanted for the current batch is
* allocated in the "batchCxt". By resetting the batchCxt at the end of
* each batch, we free all the per-batch storage reliably and without tedium.
+ * Note that data associated with the temp files lives in the "fileCxt" context
+ * which lives during the entire join as temp files might need to survives
+ * batches. These files are explicitly destroyed by calling BufFileClose()
+ * when the code is done with them.
*
* During first scan of inner relation, we get its tuples from executor.
* If nbatch > 1 then tuples that don't belong in first batch get saved
@@ -348,6 +356,7 @@ typedef struct HashJoinTableData
MemoryContext hashCxt; /* context for whole-hash-join storage */
MemoryContext batchCxt; /* context for this-batch-only storage */
+ MemoryContext fileCxt; /* context for the BufFile related storage */
/* used for dense allocation of tuples (into linked chunks) */
HashMemoryChunk chunks; /* one list for the whole batch */
diff --git a/src/include/executor/nodeHashjoin.h b/src/include/executor/nodeHashjoin.h
index d367070883..a8f9ae1989 100644
--- a/src/include/executor/nodeHashjoin.h
+++ b/src/include/executor/nodeHashjoin.h
@@ -29,6 +29,6 @@ extern void ExecHashJoinInitializeWorker(HashJoinState *state,
ParallelWorkerContext *pwcxt);
extern void ExecHashJoinSaveTuple(MinimalTuple tuple, uint32 hashvalue,
- BufFile **fileptr);
+ BufFile **fileptr, MemoryContext filecxt);
#endif /* NODEHASHJOIN_H */
--
2.39.2
--MP_/cg+MWpd=CPm0_9SBk5uZwGx--
view thread (9+ messages) latest in thread
reply
Reply instructions:
You may reply publicly to this message via plain-text email
using any one of the following methods:
* Reply to all the recipients using the --to and --cc options:
reply via email
To: [email protected]
Cc: [email protected]
Subject: Re: [PATCH] Allocate hash batches related BufFile in a dedicated context
In-Reply-To: <no-message-id-109953@localhost>
* Save the following mbox file, import it into your mail client,
and reply-to-all from there: mbox
This inbox is served by agora; see mirroring instructions
for how to clone and mirror all data and code used for this inbox