Net::NTP is consistently faster.

That is kind of a given. If you trace you Net::NTP script, it executes 1176 lines:

C:\test>perl -d:Trace junk1001.pl >junk1001.log 2>&1

C:\test>wc -l junk1001.log
   1176 junk1001.log
[download]

If you run your POE script, it executes 3575 lines:

C:\test>perl -d:Trace junk1000.pl >junk1000.log 2>&1

C:\test>wc -l junk1000.log
   3575 junk1000.log
[download]

Just over 1000 of those lines -- the ones that run in Net::NTP -- are more or less identical between the two scripts.

The other 2400 in the POE script are the overhead of running POE.

Of course, those lines enable you to do many other things besides NTP...but there is no circumstance under which you would want to have a script doing any of those other things concurrently with synchronising your system clock; because the overhead would compromise the accuracy of the synchronisation.

Actually, if you want to synchronise your clock really accurately, even the overhead of Net::NTP is too much. Especially when the whole purpose of using udp is to allow a very low latency turnaround of the request and response; and the simplicity of coding that comes from udp sendto() and recvfrom() calls.