> Overall, the gap for most AMD-based processors closed almost immediately when we controlled for NUMA nodes – in other words, when we only considered runs where each instance showed all vCPUs running across a single NUMA node. When we did this, the performance gap dropped from 22% to 1%, which is smaller than our margin of error.
How does one avoid machines with vCPUs across multiple NUMA nodes? Do you just spin the machine up, run `lscpu | grep 'NUMA node(s)'` and kill the machine if the value reported is anything but 1 and try to spin a VM again?
> Overall, the gap for most AMD-based processors closed almost immediately when we controlled for NUMA nodes – in other words, when we only considered runs where each instance showed all vCPUs running across a single NUMA node. When we did this, the performance gap dropped from 22% to 1%, which is smaller than our margin of error.
How does one avoid machines with vCPUs across multiple NUMA nodes? Do you just spin the machine up, run `lscpu | grep 'NUMA node(s)'` and kill the machine if the value reported is anything but 1 and try to spin a VM again?