Fat Datasets and Renoting

We are passing some fairly large datsets around using remoting with IIS as the host.  We have seen a lot of performance issues with doing this.  My research has revealed that a DataSet hard-codes the serialization as a diffgram regardless of the characteristics of the underlying stream.  I have heard that MS recognizes the design issue with this and is working on doing something. 

In the meantime, what are other folks doing to reduce the size of large datasets?  Any kind of compression that works?