Intel Fixing Up Sub-NUMA Clustering For Linux So That It Behaves With RDT

[phoronix]

Phoronix: Intel Fixing Up Sub-NUMA Clustering For Linux So That It Behaves With RDT

Sub-NUMA Clustering with Intel Xeon processors allows for splitting up the CPU cores, cache, and memory into multiple NUMA domains for enhancing the performance of NUMA-aware applications. While SNC can help in a number of cases especially plenty of HPC and server workloads, currently it's not properly supported if making use of Resource Director Technology (RDT) on modern Intel CPUs. That is in the process of changing with new Linux kernel patches being worked on by Intel...

https://www.phoronix.com/news/Intel-SNT-RDT-Fixing-Up

[milkylainen]

Love the picture. It's a good reminder.
LGA7259: "Oh so you think LGA4677 is big?"

[panikal]

Never heard of sub-numa clustering. This is seems to be a pretty good write-up. https://frankdenneman.nl/2022/09/21/...ma-clustering/

[ms178]

I guess this sub-NUMA-clustering is the generic term for their Cluster-on-Die (COD) featured in some Xeons?! I did some experiments with Haswell-EP and that feature the other day, but quickly found out that it is still not well-supported and saw performance regressions all over the place in games; but these are not NUMA-friendly either. At least some games that crashed before with that feature (looking at Company of Heroes 2 in particular) do work now, that's at least some progress. I hoped that the futex2-NUMA work will improve this even further, but I doubt that my Xeon will see the fruits of that work.

[S.Pam]

Never heard of sub-numa clustering. This is seems to be a pretty good write-up. https://frankdenneman.nl/2022/09/21/...ma-clustering/

I guess it is similar to AMDs LLC (Last Level Cache). Makes a massive performance impact on many workloads.

[panikal]

I guess it is similar to AMDs LLC (Last Level Cache). Makes a massive performance impact on many workloads.

According to the blog post I posted (from a principal engineer at vmware) they claim

AMD provides similar functionality called NUMA per Socket (NPS)

Though LLC does sound closer in spirit to what sub-numa clustering sounds like.