[e2e] "PMTUD using options" draft

[Michael B Greenwald]

   Thu, 12 Feb 2004 10:36:14 -0700
   Sneha Kumar Kasera <kasera at cs.utah.edu>

   In my opinion, saying that 50 percentile of the ICMP Time Exceeded 
   messages were generated in less than a millisecond is not good enough. 
   In fact I would be more interested in the higher percentiles or at least 
   the variability from the median. 

You can get that data.  In our Infocom '03 paper ("cing: Measuring
Network-Internal Delays using only existing infrastructure",
Anagnostakis, Greenwald, & Ryger,
http://www.cis.upenn.edu/~mbgreen/papers/infocom03.ps.gz (or
...pdf.gz)), Figure 2 shows our measurements of the ICMP Time Exceeded
processing time, with the 5, 25, 50, 75, 95 percentiles marked, as
well as recording all datapoints that were higher than the 95
percentile as a dot.

In fact, some of the routers probed *did* have long tails and
relatively high variability.  However, the Tulip paper in SOSP '03,
("User-level Internet Path Diagnosis", by Mahajan, Spring, Wetherall,
and Anderson) showed (Section 3.2.3)  that the long delays were
periodic (occuring every 60 seconds).  They conjecture that this is a
"coffee-break" that most likely occurs "when routers push forwarding
tables to the line cards."  I'm not sure what you want to infer from
this about slow path vs. fast path.  A millisecond is still a long
time, so the measurement granularity we are talking about here is
extremely coarse.
				    
			I think the variability would capture 
   the queuing delay in the slow path. Unless it is shown that variability 
   from the median is low or a very high percentile of ICMP Time Exceeded 
   messages were generated in less than a millisecond, I don't think one 
   could conclude that queuing delays in the fast path can be measured from 
   slow path data.
   
For most routers, the 95 percentile is around a millisecond.