I suggest that anyone having bad silo drops try recompiling their kernel
with the 16550 trigger level set to 1 character.

lines 607 and 948 should alter  FIFO_TRIGGER_8 to FIFO_TRIGGER_1.

By default it is set to 8 characters, which gives ~1400us.  Drops mean
that something in the system is eating more than 1.4 milliseconds.

Comments from people regarding the type of ethernet device and ethernet
activity might be very interesting.  It appears from isa_machdep.c that
TTY interrupts take authority over NET, and DISK.  And NET over DISK.
But I am uncertain if blocking can occur in these other devices which
makes TTY sol.

I am working on a web page where I hope i386 people can put together
their experiences, and we can try to isolate the situations which cause
com ports to be such a problem.