Batch Execution Mode And Window Functions

Kevin Feasel

2017-07-24

T-SQL

Chris Adkin shows how taking advantage of batch execution mode on rowstore tables can lead to faster performance as degree of parallelism increases:

The SQL Server execution engine fundamentally acts like a cursor, control flow is exerted from the root node of the plan down to right most child node or iterators. The (logical) flow of data through the plan is in the opposite direction.

I say ‘Logical’ because in practise the run time uses buffers in order to minimise the data movement whilst executing the plan. However, up until SQL Server 2012, which first introduced batch mode, execution plans were executed by iterators processing data in a row by agonising row manner. Batch mode changes all of this, if we can ferry rows around in batches, this reduces the number of CPU cycles it takes to process an iterator in the plan (providing it supports batch mode). Also by sizing batches such that they fit inside the level 2 cache of the CPU, we gain even more performance by minimizing CPU last level cache misses or worse still main memory.

Adkin credits Niko Neugebauer for the insight and shows how you can use this on normal rowstore tables.

Related Posts

Debugging a Pivot

Ed Elliott takes us through problems with the PIVOT statement: If you have a PIVOT query and it isn’t returning the data you expect, what can you do to troubleshoot it? The thing to do is to break it down into the constituent parts. First, lets take a look at a query and see what […]

Read More

Avoiding DONE Tokens in Loops

Emanuele Meazzo shows one reason why loops can be so much slower in T-SQL: Not everybody knows that SQL Server sends a DONE Token to the client each time that a SQL statement completes (so, everything except variable declarations); For the query above you can basically track it with extended events by tracking the “SQL Statement Completed” […]

Read More

Categories

July 2017
MTWTFSS
« Jun Aug »
 12
3456789
10111213141516
17181920212223
24252627282930
31