Slow/variable LU FFT benchmark results over network

Hello. I have a question about Matlab 2013b performance over a network.

I'm running matlab on a remote redhat E6 server with the display set back to a local redhat E6 machine (ssh -Y to connect) with KDE. The server is okay with two 8 core (16thread) Xeon E5-2660 @ 2.20GHz and 64G ram. The local is a crappy desktop (i5-3330 @ 3GHz and 8G ram) though it has a nice NVidia GTX760 4G graphics card (propriatary nvidia driver, hardware OpenGL). Connectivty is over a gigabit university network.

After testing the setup different ways I've run bench a large number of time (throwing out the first one) with some odd results. ODE and Sparse are consistenly (.1809 to .1840) and (.16 to .22) respectively which is about what I'd expect. 2D and 3D suck (1.52 to 1.77) and (1.60 to 1.69) respectively which I assume is because everything is over the network. LU and FFT results are strange. FFT results can be as fast as 0.1 and as slow as 10.5 though typical results are in the .35 to .4 range with a few 5.# and 4.# thrown in. LU is almost always 0.04 but occasionally jumps around .4 or .35.

In the past I've always sat at a good workstation so I've never seen this type of behavior. Does anyone know why the FFT would be so variable and so bad? Its slower than a 32bit 1.86GHz machine listed in the benchmarks even at its best and real crap other times. Is this network related?

Any advice about how to speed this up would be appreciated.

Thank you

p.s. forgot to mention, the poor times are not correlated as far as I can tell. here are a few I just did from 2014a:

0.043103 4.511659 0.180373 0.218745 1.457445 1.552084

0.04244 0.392074 0.180851 0.148809 1.505984 1.607132

0.291437 0.106353 0.252168 0.582083 1.458935 1.599062

0.04296 0.337787 0.184003 0.149581 1.505654 1.463433

0.042978 0.373702 0.179967 0.215923 1.446277 1.424478

jason