Use MADV_HUGEPAGE as ALLOC_TYPE_HUGEPAGES fallback

[champtar]

Subject

Use MADV_HUGEPAGE as ALLOC_TYPE_HUGEPAGES fallback

Issue type

• [ ] Bug report
• [X] Feature request

Configuration:

• Product version: VMA_VERSION: 9.7.2-1
• OS: Alma 8.6
• OFED:
• Hardware:

Actual behavior:

Using huge pages in k8s has a lot of challenges:

• you need to pick an arbitrary number of huge pages (800 ?) even if you don't know how many huge page users will run on your system
• k8s doesn't allow to over-allocate huge pages, if I have 3 pods and I give them 800 huge pages, I need my system to have 3*800 huge pages allocated before my pods start
• k8s uses cgroups hugetlb controller to limit huge pages usage, trying to use more huge pages than allocated gives you a SIGBUS (https://www.kernel.org/doc/html//v5.16/admin-guide/cgroup-v1/hugetlb.html)
• finally enabling huge pages for VMA applications means other applications might try to use huge pages and get killed with SIGBUS (k8s default limit_in_bytes is 0). An example application is pgsql (configurable but still).

When using libvma (librivermax in my case), it would be nice to fallback ALLOC_TYPE_HUGEPAGES using mmap + MADV_HUGEPAGE. Even if has no strict guarantees, with 2Mo huge pages on servers with a lot of memory it's extremely likely you will get huge pages.

Expected behavior:

libvma uses MADV_HUGEPAGE where appropriate to improve performances when explicit huge pages are complicated to setup.

Steps to reproduce: