Great write-up, and the dsync-vs-direct dd comparison is a really clean way to show that bulk throughput was fine while flushes were the problem. Two things worth adding for anyone copying this as a VM host: since mitigations=off removes the protection between guests, it's only a safe trade if every VM is your own; and on the 6,1 the single shared heatsink means sustained all-core load can throttle quietly, so logging applesmc fan speed and CPU temperature during a long sysbench run shows whether the +42% holds up over time or only in short bursts.