Metricfire logo

Base metrics

Table of Contents

  1. Load
    1. CPU utilization
    2. Load average
    3. Processes
  2. Memory
    1. Activity
    2. Virtual memory
    3. Writeback
  3. Disk
    1. iostat
    2. Capacity
  4. Network
    1. Interfaces
    2. Sockets

This document describes the base system metrics exported by the Hosted Graphite agent.

We focus on the default “base” dashboard, and also provide notes on related metrics not displayed there.

As Diamond collectors rely heavily on /proc data, many of the notes below are from Linux kernel documentation, e.g. proc.txt;

We list metric unit - percentage, count, bytes, etc. - in brackets after each metric description.

If you find anything unclear or incorrect here, please let us know!

   

Load

CPU utilization

undefined
  
User & system CPU graphs
    

These metrics are found under:

hg_agent.hostname.cpu.cpuid.*

  

and represent percentages of time each cpuid spends in particular states.

We display two of the most interesting on the dashboard:

  • user: normal processes executing in user mode (percentage);
  • system: processes executing in kernel mode (percentage).

Others you can use in your own graphs or investigations:

  • nice: niced processes executing in user mode (percentage);
  • idle: nothing to do (percentage);
  • iowait: not really reliable - see note in proc.txt;
  • irq: servicing interrupts (percentage);
  • softirq: servicing software interrupts (percentage);
  • steal: executing other virtual hosts (percentage);
  • guest: running a normal virtual guest (percentage);
  • guest_nice: running a niced virtual guest (percentage).

   

Load average

undefined
   
Load average graph

   

These metrics are found under:

hg_agent.hostname.loadavg.*

     

Load average, roughly speaking, is the average number of tasks waiting with “something to do” over a period of time:

  • 01: 1-minute load average (count);
  • 05: 5-minute load average (count);
  • 15: 15-minute load average (count).

Since the interpretation of load average is affected by the number of cores a machine has, you might like to use these “normalized” versions in your own graphs or investigations:

  • 01_normalized: 1-minute load average normalized by #cores (count);
  • 05_normalized: 5-minute load average normalized by #cores (count);
  • 15_normalized: 15-minute load average normalized by #cores (count).

  

Processes

undefined
  
Processes graph

These metrics are found under:

hg_agent.hostname.loadavg.*

These are simple “snapshot” counters of the process numbers. Note that the number running will typically be maxed out at #cores.

  • processes_total: total number of processes on the system (count);
  • processes_running: number of processes running (count).

Memory

Activity

undefined
  
Memory and swap graphs

These metrics are found under:

hg_agent.hostname.memory.*

  

In the “memory activity” graph, we display some of the metrics most relevant to physical memory usage:

  • MemTotal: total usable ram, i.e. physical ram minus a few reserved bits and the kernel binary code (bytes);
  • MemAvailable: an estimate of how much memory is available for starting new applications, requires kernel 3.14 or later (bytes);
  • Active: memory used recently, usually not reclaimed unless absolutely necessary (bytes);
  • Cached: in-memory cache for files read from the disk, i.e. the pagecache (bytes).

And “swap activity” displays:

  • SwapTotal: total amount of swap space configured (bytes);
  • SwapFree: amount of swap space available for use (bytes).

There are several other metrics available under memory.*. If you’re digging further, you can find out what they mean in the docs for /proc/meminfo.

   

Virtual memory

undefined
  
vmstat graphs
  

These metrics are found under:

hg_agent.hostname.vmstat.*

   

These are metrics from /proc/vmstat and give some insight into the activity of the Linux virtual memory system. Unfortunately, the counters are a little underdocumented.

First, pages in and out:

  • pgpgin: pages brought in from disk (count);
  • pgpgout: pages written out to disk (count).

Note that because everything goes through the page cache, these are recorded for essentially all pages read from or written to disk, so if you’re doing a lot of IO they’ll be elevated.

Next, swap usage which generally you want to keep low or nonexistent. See this article for more information.

  • pswpin: pages brought in from swap space (count);
  • pswpout: pages swapped out into swap space (count).

Finally, page faults made by the virtual memory system to page memory into process address spaces:

  • pgfault: minor page faults (count);
  • pgmajfault: major page faults (count).

Note that page faults will stimulate paging in, so you can expect these to correlate.

   

Writeback

undefined
   
Memory writeback graphs

These metrics are found under:

hg_agent.hostname.memory.*

   

  • Dirty: memory waiting to be written back to disk (bytes).

When you change disk-backed memory in the page cache, it’s not written to disk immediately, just marked as “dirty”. This graph allows you to see how much is building up & being written back over time.

   

Disk

iostat

undefined
   
iostat graphs

These metrics are found under:

hg_agent.hostname.iostat.*

   

These metrics are per-disk, and are gathered from /proc/diskstats.

  • iops: “I/O operations per second”, i.e. reads + writes (count);
  • write_byte_per_second: bytes written per second (bytes);
  • read_byte_per_second: bytes read per second (bytes);
  • util_percentage: how much of the time the disk is performing I/O operations (percentage).

There are many other iostat metrics exported per disk; you can browse your metric tree to see which and compare with /proc/diskstats and the ‘diskusage’ diamond collector.

   

Capacity

undefined
  
Disk capacity graphs

   

These metrics are found under:

hg_agent.hostname.diskspace.*

  

Again, these metrics are per-disk.

  • byte_avail: available bytes, i.e. space available for use by non-privileged users (bytes).

Apart from this useful graphed value, there are also some more available to you:

  • byte_free: available bytes for the superuser (bytes);
  • byte_percentfree: byte_free as a percentage of the total (percentage);
  • byte_used: bytes used (bytes);
  • inodes_avail: available inodes for use by non-privileged users (count);
  • inodes_free: available inodes for the superuser (count);
  • inodes_percentfree: inodes_free as a percentage of the total (percentage);
  • inodes_used: inodes used (count).
   

Network

Interfaces

undefined
  
Network interface graphs

These metrics are found under:

hg_agent.hostname.network.*

   

These metrics are per-interface. We graph the following:

  • tx_packets, rx_packets: packets transmitted, received (count);
  • tx_byte, rx_byte: bytes transmitted, received (bytes);
  • tx_drop, rx_drop: packets dropped by the driver on transmit, receive (count).

There are many other network metrics exported per interface; you can browse your metric tree to see which and compare with /proc/net/dev, which is fairly self-explanatory, and the ‘network’ diamond collector.

   

Sockets

undefined
   
Socket graphs

  

These metrics are found under:

hg_agent.hostname.sockets.*

   

They’re drawn from /proc/net/sockstat, which is under-documented.

  • used: total number of sockets in kernel socket lists (count);
  • tcp_inuse: TCP sockets currently in use (count);
  • udp_inuse: UDP sockets currently in use (count).

Others you can use in your own graphs or investigations:

  • tcp_mem: the number of pages in use for TCP (count);
  • udp_mem: the same for UDP (count);
  • tcp_alloc: number of sockets allocated for TCP (count);
  • tcp_orphan: sockets not associated to file descriptors (count);
  • tcp_tw: sockets in TIME_WAIT, i.e. waiting after close to handle packets still in the network (count).