ka-deploy

Langue: en

Version: 17 December 2001 (mandriva - 22/10/07)

Section: 1 (Commandes utilisateur)

NAME

ka-deploy, ka-d-server, ka-d-client - send large amounts of data to several machines

SYNOPSIS

ka-d-server -n NUMBER [-l] [-s SESSION] [-d DELAY] [-e COMMAND]
ka-d-client [-s SESSION | -h HOSTNAME] [-e COMMAND] [-g]

DESCRIPTION

Ka-deploy is a tool designed to send large amounts of data from one to many machines using TCP trees, typically a regular chain. The typical usage for ka-deploy is partition cloning.

The machine you want to send the data from must be running ka-d-server and the machines you want to send the data to must run ka-d-client.

The server ka-d-server will listen on a TCP port and wait for a given number of machines ( ka-d-client) to connect to him. Then the server will coordinate the clients in the creation of a TCP tree or chain. Once ready, the server will run its command and pipe its standard output into the tree. The clients will read data from the tree and pipe them into the standard input of their command.

FINDING THE SERVER

When ka-d-client is run, it must contact a ka-d-server program. There are two ways of telling the client how to find the server. The -h option can be used to give the server hostname to the client. Another way is to use sessions: give a same -s option to both the server and the client, and the client will send UDP broadcast packets to locate the server. Of course this works only if the client and server are within the same subnet. A session name is merely a short string, such as "foobar" or "helloworld".

NUMBER OF CLIENTS

The server must be given a -n option with the number of clients to wait before the building of the tree/chain occurs. If for some reason (node failure for instance) at some moment you don't want to wait for all the clients, you can force the server to exit the 'wait mode'. To do this, just run a ka-d-client with the -g option. This client will contact the server, tell him to stop waiting, and will exit.

SERVER OPTIONS

-l
Be sure the server data buffer always has enough data for recovery if a node fails. For this, the last node of the chain will report from time to time its position in the data flow to the server. I'm not too sure but this can lead to a severe performance drop with many nodes (200+).
-a NUM
Specify the 'arity' of the tree, the number of nodes under each node in the tree. The default is 1, and the result is a chain, instead of a tree.
-d NUM
Normally the server will send an 'exit signal' to the clients when they have all received the data. If this option is enabled, the server will introduce a delay between the release of the clients. This delay is the given number in tenths of second. Note that you MUST use the -w option on the clients for this to work. We use this option to smooth the load on our TFTP servers for whole-cluster system installations.

CLIENT OPTIONS

-w
If this option is used, then the client will wait for an 'exit signal' from the server before exiting. This option can be used to synchronize a bit the clients before they exit. See the -d option on the server.