Home / server / Questions / 23465
Accepted
SQLRockstar
Asked: 2009-06-11 08:25:57 +0800 CST

soft-NUMA testing

  • 772

other than staring at a clock on the wall, is there any good technique for measuring the effects of various soft-NUMA configurations for SQL?

What we are hoping to show is the performance edge gained (if any), but I was hoping to show more details than just "it used to take an hour, now it takes 48 minutes", because that is not enough proof (in my opinion) that our NUMA changes were the reason for the gain.

If there is a way to measure and decipher some output I would be most interested in hearing about them, thanks.

sql-server sql-server-2008

1 Answers

  1. Best Answer

    There's nothing definitive you can track in 2005. Maybe on 2005 you could look at the CPU utilization and show that the load is now balanced across your soft nodes, or that your network traffic is coming to multiple IP addresses, if you configured port binding (see Slava's blog at Configuring SQL Server 2005 for Soft NUMA).

    In 2008 you can use the sys.dm_os_nodes DMV and look at the active_work_count and avg_load_balance columns to see if your configuration has changed the (rough) work balance across the CPUs.

    Really its like any other kind of performance testing. If you have a test harness that produces identical conditions for each performance run, and you get consistent performance over multiple perf runs with one configuration, and then you make a configuration change and then different (and consistent) performance, then it must be the configuration change you made that caused the performance difference.

    Hope this helps.

    • 1