Arbitrage between chains and block quantities

Dear All,
Dear Johannes,
I have a little question that seems useless but very important for large models or precision purposes.

What is preferable:

  1. having a lot of mh_nblocks (5 instead of 3) and few mh_replic (3000000 instead of 5000000) ?
  2. having a few mh_nblocks (3 instead of 5) and a lot of mh_replic (5000000 instead of 3000000) ?

More generally, we are sometimes, for time and processors constraints, obliged to do a choice. My question links this choice with the quality of the results.

In other words, which configuration (1 or 2) is preferable in order to obtain best results (and why) ?

Thank you in advance,

I would usually go for more draws and fewer chains, because according to my experience convergence takes time and is less likely with fewer chains. But already with two chains you should be able to judge whether the chains have converged to the same ergodic distribution.