Traditionally, the Monte Carlo criticality calculation must set a maximum inactive step by experience to ensure that a fission source distribution has converged. The tallying process can only be invoked after this maximum inactive step to avoid the system error caused by the non-converged fission source distribution. In the same way, the uniform fission site algorithm for increasing the whole efficiency of global tallying should also be invoked after the fission source distribution has converged fully. The calculation must reach a maximum iteration step, then, this process can be stopped and the tallies can be printed. This old strategy has two defects. Firstly, the appointed maximum inactive step can only be set by experience, which will be insufficient in some cases; secondly, some iteration steps can be wasted because the precision of tallies has been enough and no one knows it. So, a new strategy is proposed in this article to overcome these defects. Based on an on-the-fly diagnostic method for the convergence of Shannon entropy sequence corresponding to the fission source distribution of each iteration step, the uniform fission site algorithm will be invoked after the iteration step whose serial number is the maximum of the first active step and the first converged step diagnosed by the above-mentioned rule. This rule will be helpful in ensuring that the uniform fission site algorithm can use enough accurate data to bias the secondary fission neutron number, thus avoiding the system error to some degree. Then, a global precision index will be calculated at each fixed step to judge whether the precision standard is reached. If so, the whole calculation is stopped. This process will be repeated until the pre-set maximum step number is reached. In this way, superfluous calculations can be skipped but the calculation precision can be guaranteed. In a word, this new strategy is beneficial to increasing the efficiency of global tallying in the Monte Carlo criticality calculation when appropriate parameters are adopted. This conclusion can be proved by the numerical result from the C5G7 benchmark model.