In this Benchmark Rewind I rerun and explain my Proxmox 9.0.4 tests on Intel 12th Gen (4 vCPU / 8 GB RAM / 64 GB disk VMs): RHEL 10, Fedora 42, Debian 13, Debian 12, and Devuan 6.
RHEL 10 dominated HPL, ChaCha20-Poly1305 (Botan), and stress-ng IPC (semaphores, context switching, message passing). Debian 13 landed last.
The punchline: server-mode tuning and ISA exposure decide these microbenchmarks. Things like tuned profiles (EPP/performance), BLAS choice, CPU model (host-passthrough), THP, and mitigations alignment can swing results by 2–3×.
I show the exact knobs, why systemd ≥257 changes governor control, and why a non-systemd server guest (e.g., Devuan) can make apples-to-apples benchmarking easier. If you care about real performance, pin the policy first—then compare.
Test Bed
• Proxmox VE 9.0.4 on Intel 12th Gen (host-passthrough CPU)
• Guests: RHEL 10, Fedora 42, Debian 13, Debian 12, Devuan 6
• VM shape: 4 vCPU, 8 GB RAM, 64 GB disk each
• Workloads: HPL (BLAS-bound), Botan ChaCha20-Poly1305, stress-ng (semaphore, cswitch, msg)
Key Findings
• RHEL 10 wins mostly from performance-biased defaults (tuned / EPP=performance, THP, glibc IFUNC paths) plus full ISA exposure.
• Debian 13 trails when left in balanced mode and/or using slower BLAS.
• IPC microbenchmarks are hypersensitive to frequency policy and futex/semaphore paths.
• Level the field: set EPP/performance (or use non-systemd), standardize BLAS/HPL.dat, align mitigations and THP, confirm host-passthrough.
Chapters
00:00 - Intro
01:24 - Benchmark Walkthrough
22:47 - Final Thoughts and Wrap up