In this post, we’ll focus on a brief check-list for achieving high speed data inserts into a SQL Server target table. There are many disparate sources, but I’ve yet to find a single checklist of everything you need to consider for achieving minimally logged insert operations.
The prerequisite for achieving data loads at speeds comparable to a file copy ultimately come down to two main things:
1. Insert operations need to be minimally logged.
2. Inserted data needs to be sorted according to the target table’s clustered index (primary key).
When insert operations are minimally logged, it means that the transaction log is bypassed during the insert operation – i.e. only the MDF (data) file is written to, instead of the MDF and LDF file/s. When the inserted data is sorted according to the target table’s primary key, it means that SQL server does not need to use TempDB to sort and build the clustered index. With heavy read/write operations to the transaction log and TempDB out of the way, the insert operation becomes extremely fast.
So… enough of the theory… here’s the full check-list for achieving high performance bulk inserts:
Target Table Checklist:
1. Database recovery model is either BULK LOGGED or SIMPLE.
2. Account used to connect to SQL server is effectively granted the BULK OPERATIONS privilege on the server.
3. Target table is empty.
4. Ideally, non-clustered indexes are disabled.
1. Pipeline data is ordered according to table’s clustered index.
2. SQL Destination (faster) or OLEDB Destination (more flexible) component is used.
3. Table Lock is checked on destination component
4. Names of clustered index columns is provided in the Order box (BulkInsertTabLock = True)
5. Complete entire load in a single operation (MaxInsertCommitSize = 0).
It isn’t always possible to meet all of the above criterion. In particular, if the target table is not empty, or the data stream cannot be sorted according to the clustered index, then it’s actually better not to attempt a single-operation bulk insert. This is because when the table is not empty, or when the data is not sorted according to the primary key, it means that SQL server can only begin to process the actual insert operation after the last row has been sent – which means that all of the data will end up in TempDB anyway. When this is the case, set MaxInsertCommitSize to a reasonably small number (thumb-suck around 10,000) to allow SQL to process the inserts in batches during the data streaming operation. This improves parallelism and reduces the size by which TempDB will grow to accommodate the operation.
1. If running SSIS on the target server, use the SQL destination component – it’s up to 15% faster. Be sure to execute the package under a user account with sufficient privillages to open shared memory with SQL Server.
2. Whenever possible, lock the table during inserts.
3. If you’re inserting sorted data into an empty table, then:
- For the SQL Destination, ensure:
- MaxInsertCommitSize = 0
- BulkInsertTabLock = TRUE
- BulkInsertOrder = columnname [, columnname, ...]
- For the OLEDB Destination, ensure:
- FastLoadMaxInsertCommitSize = 0 (or it's default - 2147483647)
- FastLoadOptions = TABLOCK,ORDER(columnname ASC)
To set up ordering for the OLE DB Destination component, edit FastLoadOptions property. It contains a comma seperated list of text, more details of which are available at http://msdn.microsoft.com/en-us/library/ms141237.aspx and http://msdn.microsoft.com/en-us/library/ms177468.aspx.