readme.pastry

这是一个著名的应用层组播中间件的源码

README.pastry

Contained in this file are notes on Duke's MACEDON implementations of
Pastry, Scribe, and SplitStream.  There are two versions each of Scribe and
SplitStream: the ones marked 'ms' are the implementations as we understood
them to be at the time of the SOSP conference, and the SplitStream papers
published therein, and the others reflect changes described to us by Ansley
Post, which we understand to be revisions to the original SplitStream
protocol.

The basic differences between sosp and non-sosp versions:
- Instead of a spare capacity group, anycast joins are sent through the 
  actual stripe tree.
- Instead of keeping a nodes visited list on the anycast messages, a 
  nodes-to-visit list is maintained.
- Scribe heartbeats were removed, and Scribe liveness was used instead to
  detect failures in the multicast tree (see problem 2)

Known Issues and Questions:
1. There is some challenge dealing with the source of the information.  
  On whether it should participate in Scribe join or not.  In the current 
  implementation -- the Scribe sender automatically ignores (forwards) 
  join messages unless the sender also happens to be the root of the
  tree.
2. The removal of Scribe heartbeats makes it possible for network delays
  or dropped packets to cause multiple nodes to believe they are the 
  parent for a particular child.  This causes bandwidth waste, and in 
  bw constrained networks causes data delivery to be hindered.
3. There is the challenge of a node determining an appropriate number of 
  children to allow, both in SS and in Scribe.  Allowing more than the 
  link allows can be problematic.  What is a good way to set this
  in practice?
4. Our current implementation uses 32 bit keys, not 160 bit keys.
5. How best to implement Scribe's "routeIP", that is -- sending directly to 
  the pastry node the hash resolves to.  Our implementation uses a simple
  cache to manage id->ip mappings;
6. The Scribe paper says that on receiving a pushdown, the peers are probed 
  for latency.  Out implementation does not do this on pushdowns.
7. Scribe path compression is not implemented
8. An optimization in the SS paper (only keeping the path to root when needed)
  is not implemented.  Can you describe this in more detail?
  
We are very interested in discussing with you differences which you notice between 
our implementations and your own.  Please contact us at [e-mail],
with questions.