We consider the uplink of multicell multiple-input-multiple-output systems, where the base stations (BSs), which are equipped with massive arrays, simultaneously serve several terminals in the same frequency band. We assume that the BS estimates the channel from uplink training and then uses the maximum ratio combining technique to detect the signals transmitted from all terminals in its own cell. We propose an optimal resource allocation scheme that jointly selects the training duration, the training signal power, and the data signal power in order to maximize the sum spectral efficiency for a given total energy budget spent in a coherence interval. Numerical results verify the benefits of the optimal resource allocation scheme. Furthermore, we show that more training signal power should be used at low signal-to-noise ratios (SNRs) and vice versa at high SNRs. Interestingly, for the entire SNR regime, the optimal training duration is equal to the number of terminals.