Can you set up a remote kernel for computation and a local kernel for predictive interface?

I am able to launch remote kernels from my Windows box to our Linux cluster that

1. has the data I need and
2. the capacity to run analysis (numerous & speedy cores, spacious memory, etc.).

Works great except for the fact that there's a bit of latency. That's not a huge deal given the cost of the calculations I run, but what kills me is the predictive interface. Any time Mathematica tries to be helpful with the predictive interface functions, the latency kicks in, the program stops responding, and I have to wait for the remote kernel to return its results before any part of Mathematica becomes responsive.

Is there a way to peg the local kernel to user interface or predictive interface tasks and assign the remote kernel(s) to computation only?

-
You can turn off the predictive interface, which might be the most practical solution here. Out of curiosity, do you have 9.0.0 or 9.0.1? –  Szabolcs Nov 25 '13 at 18:40
You could also use the parallel tools and run the subkernels on the remote machine. This way you will be able to manage (ParallelEvaluate) what computations run remotely what what runs locally. –  Szabolcs Nov 25 '13 at 18:41
Szabolcs - thanks for the practical solution: I've done that in the mean time and it helps tremendously. I happen to be running 9.0.0, though I have the installation binary to 9.0.1 available--just haven't gotten around to installing it yet. Is there a significant difference between the two? –  user2946082 Nov 25 '13 at 19:28
I'm not sure, but some predictive-interface-related (and other) bugs were supposed to be fixed in 9.0.1, I would definitely install it, and try if it's better. It's just a guess. There's also this thing, which suggests that your problem might eventually go away (in a future version) (??). Regarding the ParallelEvluate-based solution, one inconvenience is that ParallelEvaluate doesn't automatically DsitributeDefinitions like e.g. Parallelize does. –  Szabolcs Nov 25 '13 at 19:31
This is an interesting idea, but even if you could set this up, the first thing the predictive interface would need to do is pull your output from the remote kernel back to the local for analysis. Depends what the latency hangup is really attributed to. –  Joel Klein Nov 26 '13 at 19:20