Files in large storage systems are either replicated or coded at multiple servers to protect against server failures. Load balancing can be used to exploit this redundancy in storage to reduce mean file-access delay. It is easy to show that at light loads (i.e., when the rate of file access requests is small), coding leads to smaller mean file-access delay than replication. In this talk, we will use mean-field analysis to show that, for a given storage capacity per file, coding strictly outperforms replication at all traffic loads. Joint work with Bin Li and Aditya Ramamoorthy.