Distributor Features

Some of Distributor's more unique or advanced features:

Service Testing

Distributor includes built-in support for testing the back end servers. The basic operation of a load balancer is to attempt to make a TCP connection to each back end server in turn until the load balancer finds a server it can connect to. This is done for each incoming client connection. This is fine for detecting gross failures like servers that are off, etc. but is insufficient for detecting more subtle failures like overloaded servers that accept a TCP connection but then immediately disconnect, or perhaps worse accept the TCP connection but then never transmit any data or disconnect. Or servers that are serving the wrong data for some reason.

To handle these types of situations, the load balancer needs to regularly scan the back end servers using a test that simulates a client request and ensure that each of the servers is serving valid data. Any server that isn't doing so should cease to receive new connections until it is corrected.

Distributor includes just such a mechanism. This testing can be as simple or as complicated as you need. Distributor uses a plugin mechanism to make it as simple as possible to develop your own tests. The tests can either be developed as a Java class or an external application in any language you want (shell script, Perl, C, etc.) Distributor comes with LDAP and HTTP service tests.

Client Disconnection

One possible failure mode mentioned earlier is a server that accepts a TCP connection but then never transmits data or disconnects. This could leave clients stuck in a bad situation. In some cases you may wish that the load balancer would terminate all of the connections to a server when it detects that the server is not serving valid data. This might allow the clients to reconnect to a functioning server, for example. Distributor allows you to enable this behavior if you desire.

Online Control

Distributor supports online (i.e. while it is running) control of certain configurables (enabling/disabling servers, log levels, etc.) as well as statistics reports. Access is via a TCP port available only to localhost on the machine running the load balancer[1]. Interactive control can be accomplished via telnet and scripting is fairly easy using something like netcat.

[1] No, this isn't the most advanced security in the world. However, you probably shouldn't be using a box with untrusted users as a load balancer. Code to implement some form of authentication would be welcome.

Grouping of Servers

You can group servers into multiple groups. Distributor will distribute connections to servers in the first group unless it detects that all of the servers in that group are unavailable, then it will rollover to the servers in the next group, etc. This allows you to have backup servers that don't get traffic until your primary servers are down.

Connection Distribution Methods

Distributor offers two algorithms for distributing connections to back end servers. The default is round robin, distributor maintains a list of the servers and just picks the next one in the list for each connection. This should maintain a fairly even load on the servers.

If your service requires that clients always get connected to the same back end server you can choose the hash algorithm. In this case distributor maintains a hash table of client IP addresses and the server they were first connected to. When a new connection from a client comes in distributor first tries to connect the client to the same server it was connected to previously. If that server is down, or if the client hasn't connected before, distributor falls back to the round robin algorithm.

SourceForge.net Logo

$Id: features.html,v 1.3 2003/06/24 23:14:03 jheiss Exp $