I've rambled about it somewhat incoherently at sql-sasquatch.blogspot.com. (sql_handle, lonnyniederstadt, sql_sasquatch are all me - there's an idea behind my multiple identites that I'm notmanaging too well yet across multiple devices 🙂 Folks might not come back to my blog to read more since they haven't been able to figure out where I'm going yet - kind of stream of consciousness until I've got the spreadsheets and graphs to show. But what I am seeing in SQL Server basically aligns with what Oracle must have seen somewhere along the line before Oracle 11g. They disable database NUMA support by default on all platforms 🙂

Trace flag 8015 makes SQL Server ignore the NUMA node boundaries for scheduler groups and memory management. The server effectively becomes one big NUMA node... or like one big VM without pass-through NUMA.

My world is all about optimizing a heavy ETL that has more queries than logical processors, and a batch report workload with more queries than logical processors. ETL and batched reports both hit disk hard.

With repeatable and repeated performance test loads, trace flag 8015 + 8048 led to approximately 25% reduction in disk IO and 10% reduction in elapsed time even when canceling out SAN cache effect, etc. Statistics were never updated between perf tests, plans didn't change... PLE in each NUMA node was almost balanced with NUMA enabled. Our primary workloads just work better with a single memory resource and a single scheduler group, on both SQL Server 2008R2 and SQL Server 2012.