So we want some application level namespace and a system to translate
it to addresses. What characterstics should such a system have? It
needs to be highly available, so you need replication among multiple
servers. You can't concentrate all the data in one place so it needs
to be distributed, perhaps a good way of doing this would be to make
it hierarhical. It would be good if short queries and answers could
be sent using UDP to avoid the setup and teardown overhead of TCP. It
could be a directory system with complex queries but it really only
needs to be a name system that works by exact match. And it should be
secure, which tends to make things more complex and slower, but will
frequently be necessary.
So, what are you suggesting other than re-inventing DNS and/or getting
DNS security actually deployed? If you came up with something other
than DNS, why would it have any better performance? How are you going
to stop some people from designating too few servers subject to too
many common mode failure mechanisms on inadequate machines with
inadequate connectivity regardless of the naming system your run?
Donald
From: Christian Huitema <[EMAIL PROTECTED]>
Message-Id: <[EMAIL PROTECTED]>
Date: Wed, 01 Dec 1999 11:43:47 -0500
References: <Your message of "Tue, 30 Nov 1999 19:42:53 GMT." <3.0.32.19
[EMAIL PROTECTED]>
>At 10:49 PM 11/30/99 -0500, Keith Moore wrote:
>
>>note also that DNS is often slow, and seems less reliable than IP.
>>by increasing the reliance on DNS you increase the probability of failure.
>
>Data point: out of 40,000 random DNS requests logged on my work station
>over the last year, 20% underwent at least one retransmission, resulting in
>service times larger than 2 seconds. The average packet loss rate on the
>regular IP service only explain about half of these retransmissions, which
>makes me suspect that a lot of additional losses are caused by congested
>DNS servers. Increasing our reliance on the DNS is definitely not a good idea.
>-- Christian Huitema
