It's probably not an SMP issue per se, but has to do with the total
number
of running processes.
UltraSPARC processors have 4K context IDs, one of which is required
for
the pmap of each active process (well, 4095 actually since 0 is
reserved
for the kernel). When you have more than 4K processes, and the kernel
wants to execute a new process, a context ID must be stolen from
another
process and that context flushed from the MMU. The code that does the
context stealing has gotten much more complicated over the years and
now seems to be tickling some sort of bug.