[netperf-talk] Paced request/response should now report a reasonable rtt

[Rick Jones]

Folks -

I just checked-in some changes to the top-of-trunk where netperf now keeps track 
of how long it sleeps when it pauses for the --enable-intervals pacing.  This 
allows it to report a much more plausible round trip latency for a request 
response test.  The "RT_LATENCY" output now is much closer to what is reported 
for MEAN_LATENCY when one uses the global -j option to get enhanced statistics 
(which keeps an enlarged histogram of individual RTTs).

I make the handwaving assumption that the additional CPU overhead of the pair of 
timestamps was worthwhile, but it will only happen if one is *actually* pacing 
the test.  A subsequent enhancement may be to make that when one is actually 
pacing *and* the test is request/response.

If one has enabled spinning for finer granularity on the intervals, the overhead 
is nil since netperf would already be spinning on timestamp calls :)

The changes are in the "omni" code, so you can see the effect with an omni test, 
or a migrated "classic" test.

happy benchmarking,

rick jones