Re: [Blocks-development] the GRID proposal

[Blat F. <pet...@ho...>]

>From: "Ben Houston" <be...@ex...>
>To: <blo...@li...>
>Subject: [Blocks-development] the GRID proposal
>Date: Tue, 15 Aug 2000 20:35:52 -0400
>
>Hi,
>
>A few people (Jonathan Byron and Michael Spencer) have mentioned in the 
>last
>days the BLOCKS project.  I thought that I would drop by and tell you about
>my self organizing protocol that I am developing:
>
>Check it out at: http://www.exocortex.org/grid.html
>
>And yesterday I developed the start of a growing preference algorithm in
>order to grow it in a manner that will maintain coherence:
>http://www.exocortex.org/growing_v01.ppt
>
>Although I always talk in two dimensions (planar) all the algorithms are
>scalable to any dimension greater than one.
>
>I'm working full time right now so I do not have time to polish these 
>ideas.
>Also we are in a very fast moving field (distributed file sharing,
>distributed applications) so I want to get them out as quickly as possible.

Thanks for the input. I dont know much about GNUtella other than
that it broadcasts searches. I would have thought that it does
a similar sort of routing as Blocks does with file advertisements
(searches are not transmitted in Blocknets).

Each Blocks server has upto 16 connections to other servers, and
when a new file is uploaded, it is given a unique ID (which is
placed in a blocklist) an advertisement is send to each
connection. If the advert ID is not blacklisted it is forwarded
to all connections (apart from the one it came from), and so
on for upto 6 hops (its a bit like the old flood fill algorithm).

The interconnection logic in blocks is currently quite trivial,
when a new server connects to the Blocknet, the server it connects
to advertises the new server to all connections. Other servers
then choose to connect to the new server depending on how many
connections they have, and a PRNG.

The formulae works a bit like this...

connect to new server if (16-N)*10 > rand()%100
where N is the number of current connections.

so, servers with < 7 connections always react, while servers
with 14 connections have a 20% chance of trying to connect.
The new server should accept on a first-come-first-served
basis.

This hasnt been tuned but seems to work well enough for
small Blocknets. Its unclear what will happen with massive
Blocknets. Perhaps Michael & Erik's simulator will clarify this.

Blocks also relies on each 'connection' to be indistinguishable
from all the other ones so that statistical analysis of the
Blocknet cannot predict the location in the Blocknet of the
originator of published files. This makes designing Blocknet
topologies particularly difficult since servers are restricted
in the amount of information they are allowed to have about
the topology.

All comments welcome :-)

ttfn

PG.



________________________________________________________________________
Get Your Private, Free E-mail from MSN Hotmail at http://www.hotmail.com