[prev in list] [next in list] [prev in thread] [next in thread] 

List:       mojonation-devel
Subject:    Re: [Mojonation-devel] Raph's ideas on reliability
From:       "Mojo Eddie" <mojoeddie () hotmail ! com>
Date:       2001-07-03 14:53:19
[Download RAW message or body]

>I am even more impressed with you for the analysis of distributed
>reliability that you posted to #mojonation, which suggests that
>individual block servers should wholesale for blocks that match their
>hash more closely over blocks that are requested more frequently.

Assuming that economics are still in the system, blockservers will want to 
wholesale blocks such that they increase the amount of download requests 
they get from brokers (to maximize the potential for profitable contention 
for bandwidth).  Exactly what strategy maximizes that depends on the 
behavior of the brokers.

If brokers more heavily download from blockservers whose hashes match the 
blocks they're looking for, then your suggestion is correct.  But I'm not 
certain that's the most effective download strategy for brokers to follow.

Queries are fast and cheap, downloads are slow and expensive.  Furthermore, 
the more blockservers a broker queries, the more she has available to choose 
from when she gets around to downloading the block, increasing the chances 
of getting it more quickly and cheaply.  So brokers should query as many 
blockservers as possible when looking for a block; while they'll prefer the 
servers with the best hash match, they won't stop with the close ones, 
they'll query as widely as possible constrained only by the cost of queries 
(both in Mojo and in their own resources, i.e. DoQ processing, memory, etc) 
versus the marginal probability of finding another blockserver with the 
block with each additional query.

But when it comes time to download the block, there will be no preference 
given whatsoever to how good a match the server's hash is to the block.  The 
broker will want to request the block from whichever server is the least 
loaded and has the most bandwidth available.

The preference of closely matching servers in queries will tend to increase 
the amount of requests those servers get for a particular block, but only to 
the degree that different brokers send out different numbers of queries.  
For any given number of initial queries sent by brokers, all servers that 
are "that close or closer" will be equally frequently queried for the block.

Likewise, servers that are close to the block will not be favored during the 
downloading process.  Instead, the traffic will tend to be distributed 
evenly among all servers that are "that close or closer" (weighted by the 
servers' available bandwidth, of course).

Servers should want to retain blocks based on their profitability, which is 
a function of how frequently that server sees queries for the block versus 
how many other servers get queries for it and have it available to download. 
  Retention based on profitability ensures that blocks are stored on as many 
servers as their popularity demands, and thus ensures that disk space in the 
Nation is being used efficiently.

To the degree that servers retain based on hash closeness rather than 
profitability, the Nation's disk storage will be used that much less 
efficiently.

  - Mojo Eddie


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


_______________________________________________
Mojonation-devel mailing list
Mojonation-devel@lists.sourceforge.net
http://lists.sourceforge.net/lists/listinfo/mojonation-devel

[prev in list] [next in list] [prev in thread] [next in thread]