Another data point: looking at things with top, and enabling the showing of individual threads, it's the system thread named "ioflush" that gets credited with the hung time. (That is, after each hang, it immediately is shown at the top of the list, with a large percentage of the CPU.)
Also, it seems that sudden, large writes will reliably cause hangs, but a more continuous stream of smaller ones tend to flow better -- although this is just me trying to infer stuff from my so far limited observation. -tih -- Popularity is the hallmark of mediocrity. --Niles Crane, "Frasier"
