I can't help wondering how a couple thousand context switches persecond would affect the attempt to load disk info into the L1 andL2 caches. That's pretty much the low end of what I see when theserver is under any significant load.