Just over a year ago I blogged about the enhancements that were made to the sqlserver.page_split Event in SQL Server 2012 to make it easier to identify what the splitting object was and the type of split that was being performed. Sadly what I discovered writing that post was that even with the extra information about the split type, the event didn’t give you enough information to really focus on the problematic splits that lead to fragmentation and page density issues in the database. I didn’t do a whole lot with this again until recently when a question was posted by Ami Levin (Blog | Twitter) on the MVP email list that commented that the page_split event was broken in SQL Server 2012 based on a presentation he’d seen by Guy Glantser (Blog | Twitter).
Let me start off by saying, the event isn’t broken, it tracks page splits, but it doesn’t differentiate between an end page split that occurs for an ever increasing index, versus a mid-page split for a random index that leads to fragmentation and page density issues in the database. Both of these are technically splits inside the storage engine, even if we as DBA’s don’t really care about the end-page split for a increasing key value like an IDENTITY column in the database. I had Ami pass my information along to the presenter and we traded a few emails on the subject of tracking splits with the specific focus on trying to pull out the mid-page, fragmenting splits. While going through things for the third time, it dawned on me that this is incredibly simple, based one of the demo’s that was sent to me. Just over a year ago, I also blogged about tracking transaction log activity in SQL Server 2012 using the sqlserver.transaction_log event, which can be used to track mid-page splits in a database.
Last year when I wrote about the sqlserver.transaction_log event, there were 10 columns output by the event in CTP1, but as of RC0, the events output has changed and only 9 columns are output by the event.
SELECT oc.name, oc.type_name, oc.description FROM sys.dm_xe_packages AS p INNER JOIN sys.dm_xe_objects AS o ON p.guid = o.package_guid INNER JOIN sys.dm_xe_object_columns AS oc ON oc.object_name = o.name AND oc.object_package_guid = o.package_guid WHERE o.name = 'transaction_log' AND oc.column_type = 'data';
For the purposes of identifying the mid-page splits, we want to look at the operation column that is output by the event, which contains the specific operation being logged. In the case of a mid-page split occurring, the operation will be a LOP_DELETE_SPLIT, which marks the delete of rows from a page as a result of the split. To build our event session, we are going to need the map_key for the LOP_DELETE_SPLIT log_op map. This can be obtained from the sys.dm_xe_map_values DMV:
SELECT * FROM sys.dm_xe_map_values WHERE name = 'log_op' AND map_value = 'LOP_DELETE_SPLIT';
With the map_key value, we have a couple of ways to collect the information with our targets. We could collect everything into an event_file, but that doesn’t really make sense for this event. Instead the best target for this type of information is the histogram target which will bucket our results based on how we configure the target and tell us how frequently the event fires based on our bucketing criteria. If we don’t know anything about the server in question, we can start off with a very general event session that has a predicate on the operation only, and then aggregate the information in the histogram target based on the database_id to find the databases that have the most mid-page splits occurring in them in the instance.
-- If the Event Session exists DROP it IF EXISTS (SELECT 1 FROM sys.server_event_sessions WHERE name = 'SQLskills_TrackPageSplits') DROP EVENT SESSION [SQLskills_TrackPageSplits] ON SERVER -- Create the Event Session to track LOP_DELETE_SPLIT transaction_log operations in the server CREATE EVENT SESSION [SQLskills_TrackPageSplits] ON SERVER ADD EVENT sqlserver.transaction_log( WHERE operation = 11 -- LOP_DELETE_SPLIT ) ADD TARGET package0.histogram( SET filtering_event_name = 'sqlserver.transaction_log', source_type = 0, -- Event Column source = 'database_id'); GO -- Start the Event Session ALTER EVENT SESSION [SQLskills_TrackPageSplits] ON SERVER STATE=START; GO
This event session will allow you to track the worst splitting database on the server, and the event data can be parsed out of the histogram target. To demonstrate this, we can create a database that has tables and indexes prone to mid-page splits and run a default workload to test the event session:
USE [master]; GO -- Drop the PageSplits database if it exists IF DB_ID('PageSplits') IS NOT NULL BEGIN ALTER DATABASE PageSplits SET SINGLE_USER WITH ROLLBACK IMMEDIATE; DROP DATABASE PageSplits; END GO -- Create the database CREATE DATABASE PageSplits GO USE [PageSplits] GO -- Create a bad splitting clustered index table CREATE TABLE BadSplitsPK ( ROWID UNIQUEIDENTIFIER NOT NULL DEFAULT NEWID() PRIMARY KEY, ColVal INT NOT NULL DEFAULT (RAND()*1000), ChangeDate DATETIME2 NOT NULL DEFAULT CURRENT_TIMESTAMP); GO -- This index should mid-split based on the DEFAULT column value CREATE INDEX IX_BadSplitsPK_ColVal ON BadSplitsPK (ColVal); GO -- This index should end-split based on the DEFAULT column value CREATE INDEX IX_BadSplitsPK_ChangeDate ON BadSplitsPK (ChangeDate); GO -- Create a table with an increasing clustered index CREATE TABLE EndSplitsPK ( ROWID INT IDENTITY NOT NULL PRIMARY KEY, ColVal INT NOT NULL DEFAULT (RAND()*1000), ChangeDate DATETIME2 NOT NULL DEFAULT DATEADD(mi, RAND()*-1000, CURRENT_TIMESTAMP)); GO -- This index should mid-split based on the DEFAULT column value CREATE INDEX IX_EndSplitsPK_ChangeDate ON EndSplitsPK (ChangeDate); GO -- Insert the default values repeatedly into the tables WHILE 1=1 BEGIN INSERT INTO dbo.BadSplitsPK DEFAULT VALUES; INSERT INTO dbo.EndSplitsPK DEFAULT VALUES; WAITFOR DELAY '00:00:00.005'; END GO
If we startup this workload and allow it to run for a couple of minutes, we can then query the histogram target for our session to find the database that has the mid-page splits occurring.
-- Query the target data to identify the worst splitting database_id SELECT n.value('(value)', 'bigint') AS database_id, DB_NAME(n.value('(value)', 'bigint')) AS database_name, n.value('(@count)', 'bigint') AS split_count FROM (SELECT CAST(target_data as XML) target_data FROM sys.dm_xe_sessions AS s JOIN sys.dm_xe_session_targets t ON s.address = t.event_session_address WHERE s.name = 'SQLskills_TrackPageSplits' AND t.target_name = 'histogram' ) as tab CROSS APPLY target_data.nodes('HistogramTarget/Slot') as q(n)
With the database_id of the worst splitting database, we can then change our event session configuration to only look at this database, and then change our histogram target configuration to bucket on the alloc_unit_id so that we can then track down the worst splitting indexes in the database experiencing the worst mid-page splits.
-- Drop the Event Session so we can recreate it -- to focus on the highest splitting database DROP EVENT SESSION [SQLskills_TrackPageSplits] ON SERVER -- Create the Event Session to track LOP_DELETE_SPLIT transaction_log operations in the server CREATE EVENT SESSION [SQLskills_TrackPageSplits] ON SERVER ADD EVENT sqlserver.transaction_log( WHERE operation = 11 -- LOP_DELETE_SPLIT AND database_id = 8 -- CHANGE THIS BASED ON TOP SPLITTING DATABASE! ) ADD TARGET package0.histogram( SET filtering_event_name = 'sqlserver.transaction_log', source_type = 0, -- Event Column source = 'alloc_unit_id'); GO -- Start the Event Session Again ALTER EVENT SESSION [SQLskills_TrackPageSplits] ON SERVER STATE=START; GO
With the new event session definition, we can now rerun our problematic workload for a 2 minute period and look at the worst splitting indexes based on the alloc_unit_id’s that are in the histogram target:
-- Query Target Data to get the top splitting objects in the database: SELECT o.name AS table_name, i.name AS index_name, tab.split_count, i.fill_factor FROM ( SELECT n.value('(value)', 'bigint') AS alloc_unit_id, n.value('(@count)', 'bigint') AS split_count FROM (SELECT CAST(target_data as XML) target_data FROM sys.dm_xe_sessions AS s JOIN sys.dm_xe_session_targets t ON s.address = t.event_session_address WHERE s.name = 'SQLskills_TrackPageSplits' AND t.target_name = 'histogram' ) as tab CROSS APPLY target_data.nodes('HistogramTarget/Slot') as q(n) ) AS tab JOIN sys.allocation_units AS au ON tab.alloc_unit_id = au.allocation_unit_id JOIN sys.partitions AS p ON au.container_id = p.partition_id JOIN sys.indexes AS i ON p.object_id = i.object_id AND p.index_id = i.index_id JOIN sys.objects AS o ON p.object_id = o.object_id WHERE o.is_ms_shipped = 0;
With this information we can now go back and change our FillFactor specifications and retest/monitor the impact to determine whether we’ve had the appropriate reduction in mid-page splits to accommodate the time between our index rebuild operations:
-- Change FillFactor based on split occurences ALTER INDEX PK__BadSplit__97BD02EB726FCA55 ON BadSplitsPK REBUILD WITH (FILLFACTOR=70) ALTER INDEX IX_BadSplitsPK_ColVal ON BadSplitsPK REBUILD WITH (FILLFACTOR=70) ALTER INDEX IX_EndSplitsPK_ChangeDate ON EndSplitsPK REBUILD WITH (FILLFACTOR=80) GO -- Stop the Event Session to clear the target ALTER EVENT SESSION [SQLskills_TrackPageSplits] ON SERVER STATE=STOP; GO -- Start the Event Session Again ALTER EVENT SESSION [SQLskills_TrackPageSplits] ON SERVER STATE=START; GO
With the reset performed we can again start up our workload generation and begin monitoring the effect of the FillFactor specifications on the indexes with our code. After another 2 minute period, the following splits were noted.
With this information we can go back and again attempt to tune our FillFactor values for the worst splitting indexes and rinse/repeat until we determine the best FillFactor for each of the indexes to minimize splits. This is an incredibly powerful tool for the DBA moving into SQL Server 2012, and will definitely change how we perform index fragmentation analysis and troubleshoot problems with excessive log generation in SQL Server 2012 onwards.