From the manual (p. 15):
Whether you should use rooted or unrooted trees depends on the model, for example, on whether a molecular clock is assumed. Without the clock (clock = 0), unrooted trees should be used, such as ((1,2),3,4) or (1,2,(3,4)). With the clock or local-clock models, the trees should be rooted and these two trees are different and both are different from (((1,2),3),4). In PAML, a rooted tree has a bifurcation at the root, while an unrooted tree has a trifurcation or multifurcation at the root.
(you set the "clock or not" variable with the control file, PAML will throw a warning if you set the clock to 0 but still have a rooted tree)
This is the answer in FAQ for PAML:
For most models, the likelihood values are still correct even if you use a rooted tree, but the lengths of the two branches around the root are not stable, as only their sum is estimable. For other models, neither the likelihood nor the parameter estimates are correct. So really you should heed the message and use an unrooted tree in the analysis.