txLoadBalancer 1.1.0

txLoadBalancer - A Twisted-based TCP load balancer.

Contents

• Introduction

• Usage

• Features

• Performance

• Changes

Introduction

This is a pure python TCP load balancer. It takes inbound TCP connections and connects them to one of a number of backend servers.

txLoadBalancer is a fork of Anthony Baxter’s PythonDirector. It removed all threading and asyncore code and the admin UI with the Twisted-based analogs. It also significantly reorganized the API and provided many new features (see below).

Usage

$ twistd -noy ./bin/txlb.tac

This will use the default configuration file in ./etc/config.xml; you can edit the .tac file to point to the config you prefer. Be sure to edit the config.xml file to properly reflect your services in need of load-balancing.

To enable the admin interface, your config file must have the admin section defined, with the required fields. For an example, be sure to see ./etc/config.xml. For more details, please see the configuration information in the ./doc directory.

If you are creating your own script and don’t want to use txlb.tac, you can import the application setup functions in txlb.application.

If you want to “embed” a load-balancer in your Twisted application, see the example ./bin/txlbWeb.tac.

Features

• It is a pure-Twisted TCP loadbalancer.

• Thanks to Twisted, it’s async i/o based, so much less overhead than fork/thread based balancers.

• It has multiple scheduling algorithms (random, round robin, leastconns, weighted). If a server fails to answer, it’s removed from the pool - the client that failed to connect gets transparently failed over to a new host.

• Provides an optional builtin webserver for a built-in admin UI.

• Seperate management timer services that perform such tasks as periodically readding failed hosts to the rotation, updated on-disk config files with changes made to the running server.

• A built-in SSH server for managing (and modifying) a running load-balancer instance.

• A Twisted API for adding a load-balancing service to your Twisted application without the need to run a separate load-balancer.

• The application uses an XML-based configuration file.

Performance

(This section is currently incomplete)

Duncan’s notes from 2008, tested on a 2 CPU Sun Netra 240:

• a single apache instance on Solaris 10 * starting threads: * max threads

• 2 load-balanced apache instances on Solaris 10

• twisted.web instances on Solaris 10 (same docroot)

• 2 load-balanced twisted.web instances

• PythonDirector 1.0.0 proxies for 2 load-balanced twisted.web instances

Anthony’s original notes on performance:

• On my notebook, load balancing an apache on the same local ethernet (serving a static 18K text file) gets 155 connections per second and 2850 kbytes/s throughput (apachebench -n 2000 -c 10). Connecting directly to the apache gets 180 conns/sec and 3400kbytes/s. So unless you’re serving really really stupidly high hit rates it’s unlikely to be pythondirector causing you difficulties. (Note that 155 connections/sec is 13 million hits per day…)

• Running purely over the loopback interface to a local apache seems to max out at around 350 conns/second.

Changes

From txLoadBalancer 1.0.1 to 1.1.0

• Massive API changes: competely reorganzied the code base.

• Integrated patches from Apple’s Calendar Server project.

• A new API for creating load-balanced services within a Twisted application (without the need to run a separate load-balancingn daemon).

• Added support for live interaction with load-balancer via SSH connection to running Python interpretter (Twisted manhole).

• The ability to start listening on a new port without restaring the application.

• Added a weighted load balance scheduler.

From txLoadBalancer 0.9.1 to 1.0.1

• 100% Twisted: removed all threading and asyncore code completely.

• Significan API changes.

• Dropped the web API.

From PyDirector 1.0.0 to 1.1.1 (AKA txLoadBalancer 0.9.1)

• Added support for Twisted, providing the option for all management, admin and load-balancing to utilize the Twisted reactor, skipping threading and asycore altogether.

From PyDirector 0.0.7 to 1.0.0

• Very few, mostly this is to update the project to ‘stable’ status.

• The networking code now uses twisted if available, and falls back to asyncore.

From PyDirector 0.0.6 to 0.0.7

• You can specify a hostname of ‘*’ to the listen directive for both the scheduler and the administrative interface to mean ‘listen on all interfaces’. Considerably more obvious than ‘0.0.0.0’. Thanks to Andrew Sydelko for the idea.

• New “leastconnsrr” scheduler - this is leastconns, with a roundrobin as well. Previously, leastconns would keep the list of hosts sorted, which often meant one system got beaten up pretty badly.

• Twisted backend group selection works again.

• The client address is now passed to the scheduler’s getHost() method. This allows the creation of “sticky” schedulers, where a client is (by preference) sent to the same backend server. The factory function for schedulers will change to allow things like “roundrobin,sticky”.

From PyDirector 0.0.5 to 0.0.6

• fixed an error in the (hopefully rare) case where all backend servers are down.

• the main script uses resource.setrlimit() to boost the number of open filedescriptors (solaris has stupidly low defaults)

• when all backend servers are down, the manager thread goes into a much more aggressive mode re-adding them.

• handle comments in the config file

From PyDirector 0.0.4 to 0.0.5

• bunch of bugfixes to the logging

• re-implemented the networking code using the ‘twisted’ framework; a simple loopback test with asyncore based pydir:

  Requests per second: 107.72 Transfer rate: 2462.69 kb/s received

  the same test with twisted-based pydir:

  Requests per second: 197.90 Transfer rate: 4519.69 kb/s received

From PyDirector 0.0.3 to 0.0.4

• can now specify more than one listener for a service

• ‘client’ in the config XML is now ‘host’

• fixed a bug in leastconns and roundrobin scheduler if all backends were unavailable.

• whole lotta documentation added.

• running display in web api now shows count of total connections

• running display now has refresh and auto-refresh

• compareconf module - takes a running config and a new config and emits the web api commands needed to make the running config match the new config

• first cut at enabling https for web interface (needs m2crypto)

From PyDirector 0.0.2 to 0.0.3

• delHost hooked up

• running.xml added - XML dump of current config

• centralised logging - the various things that write logfile entries need to be made consistent, and a lot of additional logging needs to be added.

• Python2.1 compatibility fix: no socket.gaierror exception on 2.1

From PyDirector 0.0.1 to 0.0.2

• refactored web publishing (babybobo)

• package-ised and distutil-ised the code