We've been working on improvements to the server code to better handle GPU2 work units better. That new code is running on 18.104.22.168 right now and (so far) appears to be an improvement. If it looks better, we'll roll it out more broadly to all 4 GPU2 servers.
We've also fixed an issue with GPU2 WU's which was leading to very large uploads. This has greatly shrunk the GPU2 WU return size and this should help out with the load a great deal.
Finally, we're looking into network issues. We currently suspect that there may be some packet shaping on FAH WUs with some ISPs. This issue should be improved by the smaller WU uploads and the server issues, even if the ISP's don't change anything.