Purely from empirical evidence, I heavily suspect that the problem you
describe is the same as the panic I described yesterday, solvable through
stopping clustered writes and reads.

I only rarely got the panic dumps that I posted about: much of the time,
the behavior was exactly what you described.

Aside: Ted, I have scads of 2100s.  Would it be any help if I were to
ship you one?

	-- Todd, cactus@hks.net