http://blog.httpwatch.com/2015/01/16/a-simple-performance-co...
I don't think server push was used in this benchmark, it's arguably a poor mans way to achieve part of what's discussed in the referenced paper (serving early dependencies...early).