With Network Function Virtualization (NFV) generating significant interest in the network operator community, many network functions, including the LTE EPC, are being built as virtualized software appliances running on commodity hardware, as opposed to custom hardware. To provide fault tolerance and scalable performance, the virtualized network functions are typically built in a clustered architecture, with a front-end load balancer distributing incoming traffic across multiple replicas, and the replicas synchronizing shared state with each other for resilience and consistency. Our work compares several distributed designs of the LTE EPC along the axis of the frequency of state synchronization across the replicas, and quantifies the performance overhead of state synchronization for control plane and data plane operations. Using experiments with our distributed EPC prototypes, we show that synchronizing state for every message incurs a prohibitive performance penalty (over 70% reduction in throughput as compared to the case of no synchronization). On the other hand, synchronizing state at session boundaries, while providing lower fault tolerance guarantees, imposes a smaller performance penalty. We believe that our open-source prototypes and experimental results will guide future distributed EPC designs, and help operators pick the right design that is appropriate for their fault tolerance and performance requirements.