Choose the best one is all the time a big question also in the standard RT-PCR. I got reliable data by using RNU6B but you can find several...you need to test several and choose the one that fits in your model.
Similar to Pietro, I have found that RNU6B (or U6 snRNA) is commonly used for normalization of miRNA RT-PCR data, but microarray data from the paper linked below found that miR-191 and other miRNAs are more stably expressed across a diverse set of distinct tissue samples making them better candidates for normalization. Of course as mentioned before the best reference genes really depends on your materials being examined, and diseased vs healthy tissues may have significant alterations to the expression of any house keeping genes that are analyzed.
Somtetimes it is better to use two or three HKG to estimate correctly changes in gene expression. There is no one the best HKG. It depends on number of factors that may influence the choice of HKG.
For tissue: The Applied Biosystems Megaplex miRNA assay pool (Applied Biosystems, Warrington, UK) and associated RT–PCR arrays, such as Applied Biosystems Taqman Low-Density Arrays (Applied Biosystems), recommend using the average of RNU44 and RNU48 (with U6 additionally in the latest version)