Sql-server – SQL server performance and tuning

I have had similar issues with SQL Server, it is possible that your server is not optimally configured. Newer Xeons come with TurboBoost, HT, etc. that can affect server performance significantly.

For example, we have had success with; Low Latency Configuration for Dell servers

The settings will be applicable to non-Dell servers, they just might have different names.

We also improved performance by setting the windows power management profile to high performance, from Balanced. A final piece is that it is recommended to reserve up to 8GB of memory for the OS on x64 servers, the default SQL install takes all memory. You might want to try 4/8GB reservation by setting your max SQL Server memory configuration to 4/8GB less than total memory.

My recommendation would be to revert to the old server if possible. If you do not have regression/automation/load scripts available, then the best you can do is to record your system activity for 1-4 hours during a high activity period. Then setup a web server the same as production, and a client machine to run the script. Run the same activity against the new server, make the configuration change and run the same activity again. Really you would want to do much more, but it doesn't appear that it would be viable and is outside the scope of this question.

FILLFACTOR only applies when you build or rebuild an index, not during normal operation. Normal operations always try to fill the pages to 100%.

If you insert a row that has a variable width, then update the row to be longer, that row will no longer fit on the page if there isn't enough extra space to store the after-image on the same page. If there isn't enough space, this will cause a page split, which is the process that creates the necessary space.

What's a bit misleading about page splits is that there are "good splits" and "bad splits," even though the performance counter counts all of them.

• A good split is when new rows are added to the end of the index, like what happens when you do your initial batch of INSERTs. The new row doesn't fit on the last page, so the storage engine must allocate a new page and logically hook it up to the last index page. The new page will probably exist physically after the old-last index page.

• A bad split is when a page must be inserted in the middle of an index: the new page is attached to the index structure logically on both sides, but may (likely) not exist in contiguous physical order to those pages.

Fragmentation is the discrepancy between the logical and physical order, and for the most part, only the bad type of page splits cause fragmentation.

Because you're inflating the size of already-existing rows, this causes the bad type of page split, which is why you're seeing high fragmentation numbers.

It's unclear what your exact process is, and how many rows are ultimately going to end up in this table. If it's something like a one-time population, do the full population process, then turn around and rebuild the clustered index with 100% FILLFACTOR.

If this process happens continuously, you could "pre-allocate" the space by doing the following: add a dummy variable-width column to the table, populate it to the max length on INSERT, and then on the first UPDATE, set the dummy value to NULL while updating the real value. This method will probably add overhead to the logging mechanism because of all the data values flying around.

In general, though, you only need to concern yourself with fragmentation when all of these factors are true:

• The index is large enough that randomly reading the pages (extents) is too slow (choose your own definition of what "too slow" means)
• The index will be scanned
• The index pages (extents) will be physically read from disk (i.e., they are not already in the buffer pool)
• The storage subsystem is poor at handling random reads (i.e., spindle-based, and the bytes won't come out of cache)

I would recommend, however, putting in place a solid index maintenance solution to keep things more or less in check. You can search this site for recommendations with regards to that.