North American Network Operators Group

Date Prev | Date Next | Date Index | Thread Index | Author Index | Historical

RE: latency vs. packet loss

  • From: Martin, Christian
  • Date: Wed Jan 13 19:10:23 1999

The question you pose cannot be answered as succinctly as you may think.
 Various factors contribute to packet loss and latency - many of which
are completely unrelated.  For example, if you purchase ATM CBR
connectivity from NY to Australia, you will see high latency, but
(assuming you are within your trffic contract) no packet loss.  Here the
latency is due to propogation delay.  Then again, you may have GigE
connectivity across the data center floor to a switching cluster
supporting a group of application servers that cannot handle the load
being delivered by the network.  Here, there isn't much latency, but
there may be significant packet loss.

Also note that the traditional Poisson distribution that early queueing
theory was founded on is not always the best model for todays Internet.
Traffic patterns at exchange points are becoming more and more fractal
in nature, and new queuing technologies such as WRED coupled with rate
limiting technologies such as CAR, take your assumptions about available
bandwidth, switching latency, and queue/buffer depth and flush them down
the drain.

The point is, this is a difficult question to answer.  If I were to make
a VERY general assumption, however, I may be persuaded to agree with
your proposed correlation, although many would find objection to this
with evidence to substantiate their claim.

Regards,
Chris

-----Original Message-----
From: [email protected] [mailto:[email protected]]
Sent: Wednesday, January 13, 1999 4:18 PM
To: [email protected]
Subject: latency vs. packet loss



How well does latency correlate to packet loss on the Internet?  For 
example, if one were to pick one of several randomly placed sites on 
the net based on lowest latency to/from point x, what percentage of 
this time would this also yield the site with the lowest packet loss 
to/from point x?  My guess is that the correlation is high (due to 
typical buffer sizes).