The RPC code in the newest SVN version r115 is still not working right. Connections go up and down all the time with lots of idles. Maybe revert it back for the time being? I can do that by hand of course if need be but it's a bit of a hassle to maintain.
FWIW, the phatk2 kernel is exactly as fast as the r115 phatk kernel on my 6990. I measured over 24 hours with one of them on the first and one on the second GPU core of that card; there was no significant deviation in the number of shares produced by each core, less than 10 shares difference.
I think I found the source of the memory leak in the previous RPCClient. If it works in my testing then I will be switching it back.
As for phatk2, it's faster than the phatk kernel for me using my 5870s, but keep in mind that the 69xx cards use VLIW4 instead of VLIW5.