diff options
Diffstat (limited to 'docs/collect/system-metrics.md')
-rw-r--r-- | docs/collect/system-metrics.md | 62 |
1 files changed, 62 insertions, 0 deletions
diff --git a/docs/collect/system-metrics.md b/docs/collect/system-metrics.md new file mode 100644 index 00000000..daaf61d7 --- /dev/null +++ b/docs/collect/system-metrics.md @@ -0,0 +1,62 @@ +<!-- +title: "Collect system metrics with Netdata" +sidebar_label: "System metrics" +description: "Netdata collects thousands of metrics from physical and virtual systems, IoT/edge devices, and containers with zero configuration." +custom_edit_url: "https://github.com/netdata/netdata/edit/master/docs/collect/system-metrics.md" +learn_status: "Published" +learn_topic_type: "Concepts" +learn_rel_path: "Concepts" +--> + +# Collect system metrics with Netdata + +Netdata collects thousands of metrics directly from the operating systems of physical and virtual systems, IoT/edge +devices, and [containers](https://github.com/netdata/netdata/blob/master/docs/collect/container-metrics.md) with zero configuration. + +To gather system metrics, Netdata uses roughly a dozen plugins, each of which has one or more collectors for very +specific metrics exposed by the host. The system metrics Netdata users interact with most for health monitoring and +performance troubleshooting are collected and visualized by `proc.plugin`, `cgroups.plugin`, and `ebpf.plugin`. + +[**proc.plugin**](https://github.com/netdata/netdata/blob/master/collectors/proc.plugin/README.md) gathers metrics from the `/proc` and `/sys` folders in Linux +systems, along with a few other endpoints, and is responsible for the bulk of the system metrics collected and +visualized by Netdata. It collects CPU, memory, disks, load, networking, mount points, and more with zero configuration. +It even allows Netdata to monitor its own resource utilization! + +[**cgroups.plugin**](https://github.com/netdata/netdata/blob/master/collectors/cgroups.plugin/README.md) collects rich metrics about containers and virtual machines +using the virtual files under `/sys/fs/cgroup`. By reading cgroups, Netdata can instantly collect resource utilization +metrics for systemd services, all containers (Docker, LXC, LXD, Libvirt, systemd-nspawn), and more. Learn more in the +[collecting container metrics](https://github.com/netdata/netdata/blob/master/docs/collect/container-metrics.md) doc. + +[**ebpf.plugin**](https://github.com/netdata/netdata/blob/master/collectors/ebpf.plugin/README.md): Netdata's extended Berkeley Packet Filter (eBPF) collector +monitors Linux kernel-level metrics for file descriptors, virtual filesystem IO, and process management. You can use our +eBPF collector to analyze how and when a process accesses files, when it makes system calls, whether it leaks memory or +creating zombie processes, and more. + +While the above plugins and associated collectors are the most important for system metrics, there are many others. You +can find all system collectors in our [supported collectors list](https://github.com/netdata/netdata/blob/master/collectors/COLLECTORS.md#system-collectors). + +## Collect Windows system metrics + +Netdata is also capable of monitoring Windows systems. The [Windows +collector](https://github.com/netdata/go.d.plugin/blob/master/modules/windows/README.md) integrates with +[windows_exporter](https://github.com/prometheus-community/windows_exporter), a small Go-based binary that you can run +on Windows systems. The Windows collector then gathers metrics from an endpoint created by windows_exporter, for more +details see [the requirements](https://github.com/netdata/go.d.plugin/blob/master/modules/windows/README.md#requirements). + +Next, [configure](https://github.com/netdata/go.d.plugin/blob/master/modules/windows/README.md#configuration) the Windows +collector to point to the URL and port of your exposed endpoint. Restart Netdata with `sudo systemctl restart netdata`, or the [appropriate +method](https://github.com/netdata/netdata/blob/master/docs/configure/start-stop-restart.md) for your system. You'll start seeing Windows system metrics, such as CPU +utilization, memory, bandwidth per NIC, number of processes, and much more. + +For information about collecting metrics from applications _running on Windows systems_, see the [application metrics +doc](https://github.com/netdata/netdata/blob/master/docs/collect/application-metrics.md#collect-metrics-from-applications-running-on-windows). + +## What's next? + +Because there's some overlap between system metrics and [container metrics](https://github.com/netdata/netdata/blob/master/docs/collect/container-metrics.md), you +should investigate Netdata's container compatibility if you use them heavily in your infrastructure. + +If you don't use containers, skip ahead to collecting [application metrics](https://github.com/netdata/netdata/blob/master/docs/collect/application-metrics.md) with +Netdata. + + |