Re: source_ping

Alex Rousskov wrote:

> NetDB database is small, usually less than 200KB for a big cache.

Which is a lot in some situations. I regard NetDB as data similar to DNS
data but derived rather than maintained, and as such similar ideas
should be applied on how to distribute the information in an efficient
manner.

> HTTP is certainly not an optimal netdb transport, but I am not
> sure if any optimization is actually needed. NetDB exchange is
> not a bottleneck and does not slow down anything.

This depends entirely on the location and load of the cache. If you are
a high-load cache peering with a few high load cached then this is no
bottleneck. If you are a low-load cache, then it is a problem.

> The only question is the freshness of the data. However, transport
> mechanism is virtually irrelevant to freshness and, most importantly,
> we do not know if any other exchange rate is better than the current
> one.

There is little use in always collection all NetDB information from all
peers. For many caches only a fraction is actually used, and it takes
some time for new information to enter the system if periodic full
updates are used.

I have a idea for a very simple extension, based on current ICP
implementations:

If the host is unknown (or expired) in local NetDB, then send ICMP ECHO
to the origin site and ICP SRC_RTT to all peers supporting
ICP_FLAG_SRC_RTT. Cache ICP rtt replies as netdb data from the peer, and
ICP rtt as NetDB rtt for the peer.

This distributes NetDB data on demand, allows each cache maintainer to
decide which freshness rules that should be used and gives peers a
realistic rtt penalty. And as more aggressive caching can be used
without loosing information on new hosts it may show a lower bandwidth
usage even for high load caches (for low load caches it certainly will).

ICP may be replaced with a new protocol for pure NetDB exchanges if we
want to eleminate ICP all together. The use of ICP above involves only
NetDB queries and does not really use ICP in it's original sense. Any
extra information recevied by using ICP os a bonus, but not needed from
a NetDB point of view.

> Digests are a different story.

True. And a different application with different goals/restrictions.

/Henrik
Received on Tue Jul 29 2003 - 13:15:55 MDT