Score:2

Recommended NUMA nodes per socket on dual AMD Epyc 7643 server with 1TB of RAM

cn flag

What do you suggest to set the NUMA nodes per socket to? It is factory set to NPS1, but I'm not sure if this is the most optimal value. (We use this device for massively multithreaded bioinformatic tasks)

The possible choice is: NPS0 (will attempt to interleave the 2 sockets together) NPS1, NPS2, NP3, NPS4, Auto (OS: Ubuntu 21.10, 5.13.0-40-generic kernel, 8x128 Gb of RAM)

Thanks in advance!

enter image description here

Score:4
br flag

What do you suggest to set the NUMA nodes per socket to?

The number of actual NUMA nodes on your system - it should be set to that and nothing else. You don't mention your system layout or any other details, let us know and we might be able to help.

Balázs Horváth avatar
cn flag
Finally I set it to AUTO, and listed the nodes (attached to the original post). I hope it is brings us closer to the solution. Also NUMA balancing is turned on with sudo -u root sysctl -w kernel.numa_balancing=1 Thanks!
br flag
But what is your server and CPU/s?
Balázs Horváth avatar
cn flag
OH sorry, I've just lost in benchmarking: GIGABYTE MZ72-HB0 board with 2X AMD Epyc 7643 CPUs. Meanwhile upgraded to Kernel 5.18RC1, unfortunately without any performance increase.
Balázs Horváth avatar
cn flag
More details about benchmarking: https://github.com/bbuchfink/diamond/issues/555
br flag
This is of interest - https://documentation.suse.com/sbp/all/single-html/SBP-AMD-EPYC-3-SLES15SP2/index.html#sec-epyc-architecture
mangohost

Post an answer

Most people don’t grasp that asking a lot of questions unlocks learning and improves interpersonal bonding. In Alison’s studies, for example, though people could accurately recall how many questions had been asked in their conversations, they didn’t intuit the link between questions and liking. Across four studies, in which participants were engaged in conversations themselves or read transcripts of others’ conversations, people tended not to realize that question asking would influence—or had influenced—the level of amity between the conversationalists.