This WEB page comes from the busTRACE 10.0 User's Manual. (Table of Contents)
Non-Uniform Memory Access (NUMA) is a multi-processing architecture where the memory access time is dependent on its location relative to the processor. A processor accessing memory on its same NUMA node will have faster access than if it were accessing memory on a different node. The busTRACE analyzer allows you to capture and analyze NUMA Node information for each I/O.
The NUMA Node Performance Tester is a small applet that is designed to run on a multiple NUMA node platform. The program start by performing a memory benchmark on its local NUMA node and then progressively gets further in distance. For example, here is a screenshot from a two NUMA node system:
The "Source" column indicates the NUMA node where the source memory buffer is located. The "Destination" column indicates the NUMA node where the destination memory buffer is located. The "CPU" column indicates the NUMA Node where the processor is located.
As you can see from the screen shot, when the memory buffers and processor are located on the same NUMA node (i.e. 0-0-0 and 1-1-1) the maximum memory throughput is achieved. When the source and destination buffer are both in a different node than the processor (i.e. 1-1-0 and 0-0-1), increased latency and decreased performance occurs. On the target system, we see up to a 30% performance impact.