Different numbers of subsamples will be fine. In fact, this is a reason to use the subsamples (by using the averages of the subsamples, you are telling the program that both situations have the same amount of information, which is not the case).
Your other question has no exact answer. For your current purposes, I wouldn't worry about the statistic being too low (but others might not agree). Your variables in the model (including random effects) are there because of the experimental and treatment design. Thus, you are not over-parametrizing. In other circumstances, you could get carried away in adding variables/factors to a model that are not important, driving down the chi-square/df statistic.
Now for some general comments. You are just touching the surface of generalized linear mixed models, as I am sure you know. This is a vast field, and there is much to learn (even for experienced data analysts). The advice you have received for your current problem should get you the information you need. But this is all tricky. There are nuances and multiple details that can affect the advice given and the approach you should take in modeling. There are also nuances in the interpretation of the output from GLIMMX. It is actually quite diffiult to give simple advice on your model fitting without having long discussions on your data collection and other issues -- something that would be diffiult in the Discussion forum. The wonderful thing about GLIMMIX is that one can "easily" do analyses that would have been extremely difficult (or very tedious) less than a decade ago. The challenge is that one can go in a lot of wrong directions in the analysis without a good foundation or good advice. If you plan on doing more analysis of this type, I recommend that you do some reading on the subject. The User's Guide for GLIMMIX will be a challenge for you, based on technical level of the material, but Example 1 has several things in common with your problem (although the example is for binomial and logit link). The examples are all available under the SAS/STAT/GLIMMIX User's Guide on your computer. This article from the SAS Global Forum may also be of help:
http://www2.sas.com/proceedings/forum2007/177-2007.pdf
The one example deals with Poisson counts and the log link. Neither of these examples is exactly the same as yours, so don't worry about different codes for the model fitting. If you really want to learn more, you should get the book SAS for Mixed Models, 2nd edition (Littlell et al.; 2006). I would still recommend getting statistical advice whenever possible, but I think these sources will provide some good information.
Good luck.
