Below I'm archiving a slightly modified quote from Alena Prokharchyk in this Gist as it contains useful information that I would like to reference later.
Memory resources in Kubernetes are mesured in bytes, and can be expressed as an integer with one of these suffixes: E, P, T, G, M, K - decimal suffixes, or Ei, Pi, Ti, Gi, Mi, Ki - binary > suffixes (more commonly used for memory), or omit the suffix altogether (https://kubernetes.io/docs/concepts/configuration/manage-compute-resources-container/#meaning-of-memory). Lowercase "m" > is not a recommended suffix for memory.
Cluster.requested.memory is comprised of corresponding node.requested.memory. node.requested.memory is the sum of requested.memory of all the pods scheduled on this node.
While most of the nodes have memory with binary suffixes -
memory": "2630Mi", two of them have a suffix "m":
node1.acme.com = 3731881984500m
node2.acme.com = 2757754880500mThese 2 hosts' suffixes result in incorrect unit on the cluster. Once the nodes having this unit, are corrected, the memory on the cluster should be set to the value having the right unit.
To correct node.requested.memory, you should go over all the pods scheduled on these nodes and check if any of them have requested.memory prefix set to "m" and correct it to be a binary suffix > Mi/Ki. "m" is a milli value. If during the addition one of the pods has "m" with the low value, the sum will match its precision.