[e2e] Re: crippled Internet
dennis at juniper.net
Thu Apr 26 11:15:49 PDT 2001
> About the datasets on delays:
> Fred Baker wrote:
> > But they appear to suggest that some packets experienced 500 ms
> > of delay
> Christian Huitema wrote:
> > What I see in the graphs is that the max delay is sometime very long,
> > which raises quite a few questions. How do we explain a delay of more
> > than one second between two I2 connected universities? Is some gigabit
> > router somewhere allowing its queues to reach 100 megabytes? Or is some
> > kind of a bug? And, if this is a bug, is the bug in the measurement
> > system, or in some routing component?
> > Assuming that this is a queuing delay, then we cannot deduce from the
> > small average and std that an I2 like network is fit for VoIP.
> Does anyone have an analysis of how many of the outlier delay
> values may be due to routing convergence problems (or other
> routing-related hits)? This would be a different context for
> thinking about the future of voip.
This is what I was thinking as well. The "routing convergence" effect
could even be something as low tech as routers which fail to forward
packets at full rate while making updates to their forwarding tables.
It seems to me that if one wants to deploy something over an existing
infrastructure one must
(1) measure the infrastructure to determine if the characteristics
of that infrastructure are a good match for the application;
(2) if it turns out that some characteristic of the infrastructure is a
bad match for the application, figure out what causes the problem
(3) fix that problem, then go back to (1)
For VoIP over general-purpose Internet backbone infrastructure it seems
like we're still at step (1). QOS queuing fixes one particular problem
(transient circuit overload), but it would be a waste to deploy if that
isn't the problem which is occurring.
More information about the end2end-interest