Used cached cursors for LevelDB Iterators. This gives a ~10x performance improvement in sysbench-mongodb tests.

3 files changed, 38 insertions, 23 deletions

diff --git a/api/leveldb/leveldb_wt.cc b/api/leveldb/leveldb_wt.cc
index 3116db51329..7aa6f7aaf9f 100644
--- a/api/leveldb/leveldb_wt.cc
+++ b/api/leveldb/leveldb_wt.cc
@@ -471,16 +471,11 @@ err:
 Iterator *
 DbImpl::NewIterator(const ReadOptions& options)
 {
+  /* Iterators own the cursor until they are closed. */
   OperationContext *context = GetContext(options);
-
-  /* Duplicate the normal cursor for the iterator. */
-  WT_SESSION *session = context->GetSession();
   WT_CURSOR *c = context->GetCursor();
-  WT_CURSOR *iterc;
-  /* XXX would like a fast duplicate for LSM cursors without position. */
-  int ret = session->open_cursor(session, c->uri, NULL, NULL, &iterc);
-  assert(ret == 0);
-  return new IteratorImpl(this, iterc);
+  context->SetCursor(NULL);
+  return new IteratorImpl(this, c);
 }
 
 SnapshotImpl::SnapshotImpl(DbImpl *db) :
@@ -597,8 +592,28 @@ DbImpl::ResumeCompactions()
 IteratorImpl::~IteratorImpl()
 {
   if (cursor_ != NULL) {
-    int ret = cursor_->close(cursor_);
-    assert(ret == 0);
+    OperationContext *context = db_->GetContext();
+    /*
+     * If we are in the same thread where the iterator was opened, and there is
+     * no cursor stashed there, return it.
+     */
+    if (cursor_->session == context->GetSession()) {
+#ifdef HAVE_ROCKSDB
+      if (context->GetCursor(id_) == NULL) {
+        context->SetCursor(id_, cursor_);
+        cursor_ = NULL;
+      }
+#else
+      if (context->GetCursor() == NULL) {
+        context->SetCursor(cursor_);
+        cursor_ = NULL;
+      }
+#endif
+    }
+    if (cursor_ != NULL) {
+      int ret = cursor_->close(cursor_);
+      assert(ret == 0);
+    }
   }
 }
 
diff --git a/api/leveldb/leveldb_wt.h b/api/leveldb/leveldb_wt.h
index c51ea9719a0..60390b361c7 100644
--- a/api/leveldb/leveldb_wt.h
+++ b/api/leveldb/leveldb_wt.h
@@ -135,6 +135,7 @@ public:
   }
 
   WT_CURSOR *GetCursor() { return cursor_; }
+  void SetCursor(WT_CURSOR *c) { cursor_ = c; }
 #ifdef HAVE_ROCKSDB
   WT_CURSOR *GetCursor(u_int i) {
     return (i < cursors_.size()) ? cursors_[i] : NULL;
@@ -193,7 +194,7 @@ class ColumnFamilyHandleImpl : public ColumnFamilyHandle {
 
 class IteratorImpl : public Iterator {
 public:
-  IteratorImpl(DbImpl *db, WT_CURSOR *cursor) : db_(db), cursor_(cursor), own_cursor_(true) {}
+  IteratorImpl(DbImpl *db, WT_CURSOR *cursor, uint32_t id=0) : db_(db), cursor_(cursor), id_(id) {}
   virtual ~IteratorImpl();
 
   // An iterator is either positioned at a key/value pair, or
@@ -228,7 +229,7 @@ private:
   Slice key_, value_;
   Status status_;
   bool valid_;
-  bool own_cursor_;
+  uint32_t id_;
 
   void SetError(int wiredTigerError) {
     valid_ = false;
diff --git a/api/leveldb/rocks_wt.cc b/api/leveldb/rocks_wt.cc
index 2447425851a..4d623854691 100644
--- a/api/leveldb/rocks_wt.cc
+++ b/api/leveldb/rocks_wt.cc
@@ -46,7 +46,7 @@ using leveldb::Snapshot;
 using leveldb::Status;
 
 static int
-wtrocks_get_cursor(OperationContext *context, ColumnFamilyHandle *cfhp, WT_CURSOR **cursorp)
+wtrocks_get_cursor(OperationContext *context, ColumnFamilyHandle *cfhp, WT_CURSOR **cursorp, int acquire=0)
 {
 	ColumnFamilyHandleImpl *cf =
 	    static_cast<ColumnFamilyHandleImpl *>(cfhp);
@@ -63,8 +63,10 @@ wtrocks_get_cursor(OperationContext *context, ColumnFamilyHandle *cfhp, WT_CURSO
 			fprintf(stderr, "Failed to open cursor on %s: %s\n", cf->GetURI().c_str(), wiredtiger_strerror(ret));
 			return (ret);
 		}
-		context->SetCursor(cf->GetID(), c);
-	}
+		if (!acquire)
+			context->SetCursor(cf->GetID(), c);
+	} else if (acquire)
+		context->SetCursor(cf->GetID(), NULL);
 	*cursorp = c;
 	return (0);
 }
@@ -277,15 +279,12 @@ DbImpl::NewIterator(ReadOptions const &options, ColumnFamilyHandle *cfhp)
 {
 	OperationContext *context = GetContext(options);
 
-	/* Duplicate the normal cursor for the iterator. */
 	WT_SESSION *session = context->GetSession();
-	WT_CURSOR *c, *iterc;
-	int ret = wtrocks_get_cursor(context, cfhp, &c);
-	assert(ret == 0);
-	/* XXX would like a fast duplicate for LSM cursors without position. */
-	ret = session->open_cursor(session, c->uri, NULL, NULL, &iterc);
+	WT_CURSOR *c;
+	int ret = wtrocks_get_cursor(context, cfhp, &c, 1);
 	assert(ret == 0);
-	return new IteratorImpl(this, iterc);
+	return new IteratorImpl(this, c,
+	    static_cast<ColumnFamilyHandleImpl *>(cfhp)->GetID());
 }
 
 Status
@@ -295,8 +294,8 @@ DbImpl::Put(WriteOptions const &options, ColumnFamilyHandle *cfhp, Slice const &
 	int ret = wtrocks_get_cursor(GetContext(), cfhp, &cursor);
 	if (ret != 0)
 		return WiredTigerErrorToStatus(ret);
-	WT_ITEM item;
 
+	WT_ITEM item;
 	item.data = key.data();
 	item.size = key.size();
 	cursor->set_key(cursor, &item);