Distributed shared memory (DSM) is one of the main abstraction to implement data-centric information exchanges among a set of processes. Ensuring causal consistency means all operations executed at each process will be compliant to a cause effect relation. This paper first provides an optimality criterion for a protocol P that enforces causal consistency on a DSM. This criterion addresses the number of write operations delayed by P (write delay optimality). Then we present a protocol which is optimal with respect to write delay optimality and we show how previous protocols presented in the literature are not optimal with respect to such a criterion.