This post is based entirely on empirical experience. I don't claim to know what I'm doing. Right now I'm just looking at performance.
To actually learn more about COSMO (implemented) and COSMO-RS (not implemented), read the following article by the creator of the methods: http://onlinelibrary.wiley.com/doi/10.1002/wcms.56/abstract
As always, the test job (benzene at b3lyp/g-31+g*) is very short, so the error margin is large. A major impetus for this is the execeptional performance of PCM in gaussian, and seemingly poor performance of nwchem using standard settings. When several numbers are quoted they come from multiple runs
Task energy - empty COSMO block:
0. Gas phase - ca 40 s
1. From scratch. Empty cosmo block - 79 s
2. Loaded movecs from gas phase, empty cosmo block - 48 s, 65 s, 65 s
The default is water and rsolv=0
Movecs loaded in all cases. Solvation energies in 
Task energy -- rsolv
0. rsolv=0 - [3.27 kcal/mol] - 48 s, 65 s, 65 s, 65 s
1. rsolv=0.5 - [3.01 kcal/mol] - 58 s, 58 s
2. rsolv=1 - [2.95 kcal/mol] - 57 s, 57 s, 58 s
3. rsolv=2 - [2.62 kcal/mol] - 55 s
The molecular volumes obtained are 74.5, 58.0 and 54.0 Å3, respectively, for r=0..1. My next post will talk about what this actually means, but in short, this has nothing to do with the solvent/solute cavity.
Task energy -- lineq
0. rsolv=0.5; lineq 0 - [3.01 kcal/mol] - 58 s, 58 s, 56 s
1. rsolv=0.5; lineq 1 - [3.01 kcal/mol] - 58 s
Task energy -- ificos
0. rsolv=0.5; lineq 0, ificos=0 - [3.01 kcal/mol] - 58 s, 58 s, 56 s
1. rsolv=0.5; lineq 0, ificos=1 - [3.01 kcal/mol] - 62 s
1 (one) kcal./mol = 4.184 kJ/mol -- there's thus a fairly wide range of values obtained above depending on the absolute settings.
Rsolve defines the probe used to find the solvent accessible surface -- the smaller the value, the more fine-grained and larger the apparent accessible surface. We would expect that a fairly small number is preferable for rsolv.
Ultimately, I don't see any obvious way of improving performance, other than using large values for rsolv.
An interesting feature is that the surface used by COSMO is save in a cosmo.xyz file in the runtime directory -- all that remains is working out a way of calculating the volume from this (I know it's reported in the nwchem output, but it never hurts being paranoid)