Hello,
Your situation is quite interesting from a standalone performance perspective. Below are some tips/questions to help followup on your situation..
1) I could not quite understand why CPU usage does not move beyond 12%, if there are no other system bottlenecks (e.g. network, I/O) !! Theoretically speaking, even if NUMA penalty did exist with this 2-socket hardware, at least half of the system should have been fully utilized and saturated (i.e. >= 50% CPU utilization) assuming the client workload requires server-side parallelism as you have confirmed. Can you throw some light on why this is not the case?
2) There is a command provided by libnuma library on Linux to print the inter-node distances obtained by reading the ACPI SLIT tables (i.e. numactl --hardware). Can you confirm the output seen on your system for below commands?
$ numactl --hardware
$ lscpu
3) The 'numactl' command on Linux also allows to bind physical resources of a process (i.e. threads + memory) to a particular NUMA node. So suppose you did this in the case of ADS, does it improve the CPU utilization % and the associated performance throughput?
4) I quite agree with Ashutosh K's suggestion that it is worthwhile to quickly disable NUMA configuration and repeat the workload to check if it improves the CPU and performance considerably. This will help quickly rule in (or out) the distributed memory factor, and if a noticeable difference is seen, we can then move on to memory-specific analysis using the Intel PCM/Vtune tools
Regards,
Mehul Wagle