My benchmark results vs real-world app performance

So I ran YABS on my new Contabo box and it came back with these really solid numbers and I was feeling pretty good about it and then I actually deployed my postgres workload and it was crawling and I mean crawling and I sat there staring at htop and the cores were barely moving and I went back and looked at the benchmark and realized YABS was hitting sustained turbo and my actual workload was getting bounced around by the scheduler and I started digging into the CFS settings and the cpu.shares vs cpu.quota difference and it turned out Contabo was using a different scheduler policy than what YABS assumes and so I went down this rabbit hole of schedtool and chrt and I found that if I pinned my database threads I got within 15 percent of the benchmark prediction and I guess my question is has anyone else seen this massive gap between synthetic and real world and is there a better benchm

Hey this is wild! I saw the same thing on my Ampere box at OVHcloud and the scheduler was absolutely thrashing my MariaDB threads! Have you tried running sysbench with the --thread-handling=pool-of-threads flag? That gets way closer to real contention!

Brutal gap between benchmark and reality. Which providers take XMR and actually let you pin cores without overselling into oblivion? Privacy means nothing if the performance is fake 🪙