Cluster: C3 Tools README

From Earlham CS Department
Revision as of 11:38, 3 August 2009 by Kay (talk | contribs) (New page: <pre><nowiki> C3 version 4.0: Cluster Command & Control Suite Oak Ridge National Laboratory, Oak Ridge, TN, Authors: M.Brim, R.Flanery, G.A.Geist, B.Luethke, S.L.Scott (C...)
(diff) ← Older revision | Latest revision (diff) | Newer revision → (diff)
Jump to navigation Jump to search
	C3 version 4.0:   Cluster Command & Control Suite
	 Oak Ridge National Laboratory, Oak Ridge, TN,
    Authors: M.Brim, R.Flanery, G.A.Geist, B.Luethke, S.L.Scott
	        (C) 2001 All Rights Reserved



for help or reporting a bug please e-mail luethkeb@ornl.gov or 
scottsl@ornl.gov

For installation instructions see INSTALL

for examples of scripts using the C3 tools and it's related libraries see
the contrib directory

node position in the c3.conf file is important. C3 version 3 and above allows 
the use of node ranges on the command line. The numbers used for position refer to 
the nodes position in the c3.conf file. To use an example, take the following
c3.conf file:

cluster local {
	htorc-00:node0  #head node
	node[1-64]      #compute nodes
	exclude 60
	node[128-256]
}

this cluster is made up of 192 nodes. the range of name "node[1-64]" is listed
as slots 0-63 in the list. the range of names "node[128-256] fill up the slots
64-192. As you can see the slot a node fills in a command line range does not
necessarily coincide with it's node name. It is also a reason to explicitly
specify that a node is dead (as opposed to commenting that line out), that 
way the node at slot 63 is always the machine labeled node64. There are two 
tools added to the C3 tools suite to help manage node numbers. cname is given
a name and returns its position, or slot, that it occupies. cnum takes a range
and returns what machine names are in that slot. Please see the c3.conf(5) and
c3-range(5) man pages for more details.

Version 4.0 and above allow the use of a scalable configuration. You divide your
cluster into smaller sub-clusters than execute in a parallel. For example

cluster part1 {
	node1
	node[1-10]
}
cluster part2 {
	node11
	node[11-20]
}

is a 20 node cluster. Since C3 can easily handle up to about a 64 node fanout on
most hardware the maximum recommended size is 64 64-way clusters, or 4096 nodes. 
See c3-scale(5) for full details.