You run a Gearman server (or more likely, many of them for both high-availability and load balancing), then have workers (using Gearman::Worker from the Gearman module, or libraries for other languages) register their ability to do certain functions to all of them, and then clients (using Gearman::Client, Gearman::Client::Async, etc) request work to be done from one of the Gearman servers.

The servers connect them, routing function call requests to the appropriate workers, multiplexing responses to duplicate requests as requested, etc.

More than likely, you want to use the provided gearmand wrapper script, and not use Gearman::Server directly.