It may be better to perform stratified sampling. A stratified sample is a mini-reproduction of the population. Before sampling, the population is divided into characteristics of importance for the research. For example, by main species type. Then the population is randomly sampled within each category or stratum. Random sampling has a very precise meaning in that each community has an equal probability of selection, which it may in fact not have.

Since communities may be of different sizes, perhaps you should compare composition by percentage, for example using Shannon's or Simpson's diversity indices.

See: On sampling procedures in population and community ecology, Vegetatio 83: 195-207, 1989.

I hope this helps :-) 

I am assuming that you already have the data on the communities. That is you have already done some sampling. If this is the case then you can use computer based resampling procedures from your data to determine the distribution of the test statistic you are interested in.

There is a lot of literature on this type of approach falling generally under the term of monte carlo methods, including bootstrapping and randomisation

Look at the following paper by Crowley for an introduction.

http://sciences.ucf.edu/biology/pascencio/wp-content/uploads/sites/139/2016/11/Crowley_1992.pdf

I would use a cluster analisis, using Jaccard index for dissimilarity, then I would do an ANOSIM to find how many groups can be found within the cluster.

To compare among communites of different sizes you can try using species equivalent values (Hill numbers) (Jost 2006 in Oikos).

All this can be performed using package vegan and vegeterian in CRAN, pretty easy