Described is a technology by which a relative load of network traffic handling is determined between servers of a cluster, based upon time values that correspond to the time taken by each server to handle a packet. Load may then be transferred between the servers based upon the relative load, for example by having a less-loaded server take some of the responsibility for processing incoming traffic from a more-loaded server. For example, the processing time of a server may be determined by when a receiving server receives a request packet, and when that server sends a return packet. A round trip time for a request and return communication may also be established. A logical tree of nodes representing the servers may be constructed to select pairs of servers for balancing with one another, with the selection algorithm operating recursively, in parallel, and/or repeatedly, until the cluster is balanced.