[jboss-jira] [JBoss JIRA] Issue Comment Edited: (JGRP-100) Large-scale JGroups

[Bela Ban (JIRA)]

    [ https://issues.jboss.org/browse/JGRP-100?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=12511492#comment-12511492 ] 

Bela Ban edited comment on JGRP-100 at 1/5/11 3:51 AM:
-------------------------------------------------------

When we have many nodes in a cluster, the header overhead becomes bigger. E.g. a UUID address is 17 bytes (2 longs + 1 byte). If we have 100 nodes, a View for example is ca 1700 bytes ! A Digest might be even bigger, as it contains addresses and low/high seqnos !
If we have a small MTU (like in Infiniband (2048 bytes !), then we need to add FRAG just on top of the transport.

The goal here is to make header and address overhead smaller. Headers by using IDs (shorts) for protocol names rather than strings. Addresses by using canonicalization ("ID" protocol), and replacing UUIDs with shorts.
This is https://issues.jboss.org/browse/JGRP-931. However, as JGRP-931 didn't yield any perf improvements (albeit on small clusters), I closed JGRP-931. Might be re-opened later.

      was (Author: bela at jboss.com):
    When we have many nodes in a cluster, the header overhead becomes bigger. E.g. a UUID address is 17 bytes (2 longs + 1 byte). If we have 100 nodes, a View for example is ca 1700 bytes ! A Digest might be even bigger, as it contains addresses and low/high seqnos !
If we have a small MTU (like in Infiniband (2048 bytes !), then we need to add FRAG just on top of the transport.

The goal here is to make header and address overhead smaller. Headers by using IDs (shorts) for protocol names rather than strings. Addresses by using canonicalization ("ID" protocol), and replacing UUIDs with shorts.
  
> Large-scale JGroups
> -------------------
>
>                 Key: JGRP-100
>                 URL: https://issues.jboss.org/browse/JGRP-100
>             Project: JGroups
>          Issue Type: Feature Request
>            Reporter: Bela Ban
>            Assignee: Bela Ban
>             Fix For: 3.0
>
>
> - Run JGroups on hundreds of nodes (either physical, or simulation).
> - Determine a protocol stack that can be used for large-scale execution
> - Example:
>   - Coordinator may be SPOF. If coord is hung, messages will be sent, but no new views will
>     be generated
>   - Retransmission: retransmit from anyone (not sender, otherwise we have NAK implosion)
>   - Look at PBCAST

-- 
This message is automatically generated by JIRA.
-
For more information on JIRA, see: http://www.atlassian.com/software/jira