Anyone who’s done parallel programming knows that it’s not easy. Safe access and locks need to be used to synchronize access and protect shared resources. The AS xVelocity engine (aka Vertipaq) is a multi-threaded application that accesses RAM as a shared resource. As each thread competes for a global lock for memory allocation/deallocation, scalability decreases. Because processor cores reserve memory in chunks, called cache lines, additional synchronization is required when threads access memory location that are close to each other. Before SQL Server 2016 SP1, this was a serious issue with Tabular. As we’ve learned, starting with SQL Server 2016 SP1, Microsoft switched to using the Intel Threading Building Clocks (TBB) C++ library. Specifically, Tabular now uses the TBB scalable memory allocator. Mind you that TBB is not CPU-specific extensions, so any modern Intel CPU should get these benefits. “Analysis Services SP1 uses an Intel TBB-based scalable allocator that provides separate memory pools for every core. As the number of cores increases, the system can scale almost linearly” and “The Intel TBB-based scalable allocator is also expected to help mitigate performance problems due to heap fragmentation that have been shown to occur with the Windows Heap”. Further, starting with SP1, Tabular is NUMA aware with 4-node NUMA system but I don’t know if the AAS VMs are preconfigured for NUMA.
Teo doesn’t have any firm conclusions at this point, but his initial testing looks positive.