Hey Andrew,
I've been working on compression of DMX for just such purposes, and found over one frame for one universe is pretty useless, though over a group of frames which are destined for separate universes, the resultant compression can be quite significant. I've used both RLE and DEFLATE and had varying results. RLE is great if you have lots of zero's, but with every channel at different levels (pixel mapped video for example) it's useless. DEFLATE is great when your compressing multiple universes at a time, and I could achieve about 500 chunks per second maximum throughput (core i5) bringing 16 universes down to about 3-4kb per chunk, at 30FPS =ing about 150kb/s (not constant of course). Even better is processing the DMX and sending delta values across universes rather than full frames (using a 16-bit value for channel/value with a separate upper/lower half universe boundary) - but only where the total size of the delta chunk is going to be less than the compressed chunk. As a chunk, timing of transmission back to DMX becomes less of an issue also since all universes are decompressed and available at the same time. Obviously I didn't use Art-Net nor sACN, but custom protocol on TCP which was transported over ADSL2 lines so naturally there was latency involved too, so 8 chunks were buffered before playback commenced and that seemed sufficient. 100 universes could be a different story though!!
All this is windows, closed source and commissioned work, so completely useless for you and OLA sorry, but the concept of a processed/compressed DMX transport stream between OLA nodes is up for grabs.
Cheers,