The invention relates to a system and method for load-balancing multiple servers in a communications network. Instead of using round robin or other predetermined scheme, SIP messages are forwarded to one of multiple SIP servers according to a performance score that is calculated from measured performance data. Advantageously, the disclosed system and method decreases signaling latency, improving overall communications speed. Moreover, where performance data indicates that a SIP server has failed, the performance score for the failed SIP server is zero, and the load balancer will not forward SIP messages to the failed SIP server. System uptime is also improved.