NUMA vs PostgreSQL
Wednesday, October 23 at 12:10–13:00
Most modern computers running PostgreSQL have non-uniform memory access characteristics (NUMA). I.e., accessing some memory is faster than accessing other memory. All modern server CPUs, most desktop CPUs and even many laptop CPUs are NUMA, due to the use of chiplets.
While PostgreSQL has been optimized to work better on NUMA systems, so far we have done so by optimizing memory accesses and algorithms in a general way, rather than
a) optimizing algorithms and data structures to take NUMA into account b) specifying memory locality (i.e. interleaved or local) for regions of memory
In this talk I'll present the problem, show the results of some experimental patches to improve NUMA awareness and will speculate on what we should do in PostgreSQL.
