If you’ve ever taken the time to read http://www.tpc.org/results/FDR/TPCC/hp_orca1tb_win64_fdr.pdf, you might find a bit in there that’s confusing.

 

"Microsoft SQL Server 2005 Enterprise Edition 64-bit Service Pack 1 was configured to utilize "soft NUMA", a feature that allows network connections to be affined to specific groups of CPUs (this is independent of the hardware NUMA features of the HP Integrity  Superdome). SQL Server was configured with 33 SoftNuma nodes. 31 of these nodes were configured with 2 CPUs each. The 32nd SoftNuma node as configured with one CPU and used for the checkpoint process. The 33rd SoftNuma node, also with one processor, was reserved for the Log Writer thread. A script doing continuous checkpoints of 1700 seconds (28 minutes, 20 seconds) was started on one of the server after steady state was reached with a connection port that connected to the 33rd Numa Node.  This allowed the main checkpoint process to run on that single processor, which directed the processing of the checkpoint tasks that SQL assigned to each of the other 31 SoftNuma Nodes."

 

What’s confusing about this is it’s reference to the checkpoint process, since the BOL also says that you can configure I/O affinity and CPU but not specific threads.

 

As it turns out, this issue is explained in the BOL article “How to” Map TCP/IP Ports to NUMA Nodes”.  The “connection process” mentioned above is not the more commonly know background checkpoint that flushes dirty pages to disk, but is a separate connection affinitized (is that even a word?) to the 32node which issues a CHECKPOINT command. 

 

Thanks to Eladio Rincon (Solid Quality Mentors) and Peter Byrne (Microsoft) for working this one out.