Now that there's a Parallella optimized for headless operation in the product catalogue, I think shodruk's gl_streaming is more appropriate than ever.
I'm looking forward to using this solution once RS starts stocking the Micro-Server. It represents very good factoring of functionality to let each board do what it's best at, and Epiphany will excel at computational tasks. The Micro-Server model delivers the most Epiphany for the buck, so this model of Parallella is perfect when that's one's primary interest.
Morgaine.