Neither, I had best results with fastMNN from the Bioconductor package batchelor. Depends on the magnitude of the batch effect though. Sometimes, if it is only some minor technical variation, like the basically same material prepared on two different days maybe even something simple as regression on the logcounts can be enough. On the other hand if the actual difference that you aim to remove is a pertubation effect that confounds the clustering then more stringent methods are required. You would need some details about your samples. What are they, are the same celltypes expected in both batches, is there a pertubation involved...

As usual in this context I will recommend to read the relevant paragraph in our holy single-cell bible OSCA: http://bioconductor.org/books/release/OSCA/integrating-datasets.html

You may also find this previous discussion helpful: about batch correction in scRNA-seq