exporter

perf_exporter

perf_exporter is a Prometheus exporter that exposes metrics from the perf subsystem in Linux. It can read any kernel tracepoints and expose them as Prometheus compatible metrics.

Configuration

The configuration format allows you to specific profilers at the subsytem level. For each subsytem individual events can be configured. Note that configuring a subsystem event for a specific processor isn't supported as of now. To find available events for your system you can use the perf tooling (i.e. perf list) or you can read directly from tracefs available_events in combination with the tools/tracepoint2yaml script. Here is a rough example of a configuration file to get started (note this is highly system specific).

kmem:
  events:
    - mm_page_alloc_extfrag
    - mm_page_pcpu_drain
    - mm_page_alloc_zone_locked
    - mm_page_alloc
    - mm_page_free_batched
    - mm_page_free
    - kmem_cache_free
    - kfree
    - kmem_cache_alloc_node
    - kmalloc_node
    - kmem_cache_alloc
    - kmalloc
net:
  events:
    - netif_rx_ni_entry
    - netif_rx_entry
    - netif_receive_skb_entry
    - napi_gro_receive_entry
    - napi_gro_frags_entry
    - netif_rx
    - netif_receive_skb
    - net_dev_queue
    - net_dev_xmit
    - net_dev_start_xmit

Note that the proper value for perf_event_paranoid should be set, in this case it should be set to 0 becuase the exporter runs on all processors. For more info see man perf_event_open.

Building

This repo uses make for the build system, to build the binary just type make. It is assumed that you are using go 1.11+.

Example

Here is an example of some of the events that can be exposed:

FAQ

• How is perf being used? You may want to see this library which is where most of the perf related utilities are.
• I don't see values for my perf events, is the collector broken? This is difficult to debug due to a large number of factors at play. Everything from the way your kernel was configured to debugfs mount points can cause an issue, please file an issue so that datapoints can be collected.
• Is there a max number of events that can be profiled? Yes, it is dependent on your kernel configuration, originally there was a --yolo flag to trace everything but that didn't work out so well.