I've been giving BOOTP a hammering on RiscBSD. I have 20 clients
attempting to boot simultaneously. Upon failing, they retransmit at the
interval.  This means that there are 20 bootp broadcasts at once each time
(the less said about the lack of backing off as per rfc0951 the better,
grrrr.).  Basically what happens is that bootpd takes up about 15% of CPU
time, but forces the time spent in system up to the 90% mark. Doing a
bootpd -d 10 (to force debug mode) claims replies are being sent. The
upshot of which no machines boot (or rather if they do, they seem to do it
one by one). If the machines are switched on at 5 second intervals they
all boot up fine. 

So my question is: Why does bootpd and/or the kernel cope so badly with
such a situation?

Stephen