In this work, we aim to determine an optimal strategy for accelerating high-resolution 3D myocardial T2 mapping. We quantitatively evaluate the performance of diverse methods involving different subsampling patterns and different reconstruction strategies relative to volume-by-volume SENSE reconstruction. Reconstructions which address all volumes as a single reconstruction problem (i.e. joint-sparsity SENSE or model-based SENSE) outperform volume-by-volume approaches, and variable density sampling outperforms equal spacing or CAIPIRIHNA undersampling. The T2 values observed in parametric maps proved to be more sensitive to data corruption than images themselves, limiting the degree of data reduction tolerable.