I dunno. The paper by LaMacchia and Odlysko on how to break Diffie-Hellman quickly once you've done a lot of precomputation on a static modulus is sufficiently disturbing to me that I would prefer to be able to change modulii fairly frequently if possible. Quoting K. McCurley about the above mentioned work: "Their experience seems to suggest that it is possible to compute discrete logarithms in groups GF(p)^* with p \wavyequals 10^100." [in _The Discrete Logarithm Problem_, collected in _Cryptology and Computational Number Theory_] The security of a 1000-bit modulus is just fine, thank you very much. Some military applications evidently use twice that, though. You need to change it as often as you change RSA keys. Since you can factor if you can take discrete logs, you've got to worry about the security of your RSA keys at the same time.
In addition, changing the modulus can have unpleasant effects on traffic analysis, if not done properly.
Of what sort? For D-H, the modulus must be transmitted in the clear. Unless you use a different modulus for each conversation, there is a persistency to the moduli that gives rise to a pseudo-identity. Eric