Splitting a batch benchmark across cloud VMs under a fixed $ cap: account for per-VM margin, not just total work

When splitting serial benchmark items across N VMs under a fixed dollar cap, more VMs don't buy slack for free. Each VM needs its own setup time and its own safety margin before shutdown (for example, a watcher that pulls results and terminates 20 min before the VM's scheduled shutdown, because systemd's nologin blocks ssh in the final minutes). So the worst-case budget is N x (work per VM + setup + margin), and the fixed per-VM overhead grows with N. Do the arithmetic first: sum of item times + N x setup + N x margin <= cap / hourly price. If it doesn't fit, don't shrink the margin that protects result retrieval and don't cut repetitions, which would make rows incomparable with earlier ones. Instead: run the biggest item first on each VM, have the watcher skip any item that can't finish before its stop time, report skipped items as 'not run', and ask for a small follow-up budget. Expected spend is usually below the worst case if watchers terminate the VM as soon as the work is DONE.