memory performance on dual socket server HP DL380 Gen 10

[DavorK]

Hi all,

I have some weird performance issues with dual socket dl 380 gen 10.

Memory throughput with single socket cpu is around 10gb/s.

But when i put second CPU the memory performance is around 4gb/s.

CPU is Intel(R) Xeon(R) Gold 6148 and memory is 256GB 2666MHZ. ( Every socket have 128GB RAM ( 4 modules with 32gb ))

Bios is latest and every single profile in bios dont change the throughput.

I have tested the performance via Novabench on different os ( windows / debian / xen )

Can anyone help with this?

[rmay_bk]

Novabench isn't NUMA-aware. The processes are running on both nodes (CPUs) and are accessing memory on both nodes indiscriminately.  Some of this memory access "crosses over" with CPU0 accessing CPU1's memory and vice-versa.  Latency is significantly higher in this scenario.  A NUMA-aware workload understands memory locality and attempts to minimize this activity, keeping most memory access "local" within the node.

You may be able to improve the results by adjusting the NUMA configuration in BIOS. You might get it to run full speed if you "pin" the benchmark processes to the cores in a single CPU.