On 9/6/2024 2:37 PM, Morten Brørup wrote:
From: Anatoly Burakov [mailto:[email protected]] Sent: Friday, 6 September 2024 13.47 To: [email protected] Subject: [RFC PATCH v1 0/5] Adjust wording for NUMA vs. socket ID in DPDKWhile initially, DPDK has used the term "socket ID" to refer to physical package ID, the last time DPDK read "physical_package_id" for socket ID was ~9 years ago, so it's been a while since we've actually switched over to using the term "socket" to mean "NUMA node". This wasn't a problem before, as most systems had one NUMA node per physical socket. However, in the last few years, more and more systems have multiple NUMA nodes per physical CPU socket. Since DPDK used NUMA nodes already, the transition was pretty seamless, however now we're faced with a situation when most of our documentation still uses outdated terms, and our API is ripe with references to "sockets" when in actuality we mean "NUMA nodes". This could be a source of confusion. While completely renaming all of our API's would be a huge effort, will take a long time and arguably wouldn't even be worth the API breakages (given that this mismatch between terminology and reality is implicitly understood by most people working on DPDK, and so this isn't so much of a problem in practice), we can do some tweaks around the edges and at least document this unfortunate reality. This patchset suggests the following changes: - Update rte_socket/rte_lcore documentation to refer to NUMA nodes rather than sockets - Rename internal structures' fields to better reflect this intention - Rename --socket-mem/--socket-limit flags to refer to NUMA rather than sockets - Add internal API to get physical package ID [1] The documentation is updated to refer to new EAL flags, but is otherwise left untouched, and instead the entry in "glossary" is amended to indicate that when DPDK documentation refers to "sockets", it actually means "NUMA ID's". As next steps, we could rename all API parameters to refer to NUMA ID rather than socket ID - this would not break neither API nor ABI, and instead would be a documentation change in practice. [1] This could be used to group lcores by physical package, see e.g. discussion under this patch: https://patches.dpdk.org/project/dpdk/cover/20240827151014.201-1- [email protected]/Thank you for cleaning this up, Anatoly. I would prefer to take one more step and also rename functions and parameters, e.g. rte_socket_id() -> rte_numa_id(). For backwards compatibility, macros/functions with the old names can be added.
I don't think we can do such changes without deprecation notices, but it's a good candidate for next release.
I have thought about including parameter renames in this patchset, but for now I decided against doing so. I can certainly include this in the next revision if that's something community is willing to accept.
-- Thanks, Anatoly
