Monitoring with Ganglia

Chapter 8. Ganglia and sFlow

Peter Phaal

Ganglia’s gmond agent already has built-in metrics and can be extended using plug-in modules—why do I need to know about sFlow? The short answer is that sFlow agents are available for platforms such as Windows servers and hypervisors that aren’t currently supported by gmond. A longer answer requires a basic understanding of how sFlow integrates with Ganglia to extend coverage and improve efficiency.

There are strong parallels between Ganglia’s approach to monitoring large numbers of servers and the sFlow standard used to monitor the switches connecting them. The scalability challenge of monitoring the network links mirrors the challenge of monitoring servers because each server has at least one link to the network. However, the constraints are different, leading to divergence in the functional split between generating and consuming metrics.

Network switches perform most of their functionality in hardware and have limited processing and memory resources. While computational resources are scarce, switches are richly connected to the network and excel at sending packets. With sFlow, raw metrics from the switches are sent over the network to a central server, exploiting the relatively abundant network resources to shift processing and state from the switches to software running on the server. Removing state from the switches minimizes the memory footprint and eliminates the need to dynamically allocate memory—both useful properties when embedding ...