banshee/linux-collection/linux/snmp/diskio

Disk input/output of a Linux host, by SNMP

Returns, for each disk, the reads and writes per second, their throughput, and the share of time the disk is busy. Two watches are offered, each adjustable on the probe.

How it works

The collector reads diskIOTable (UCD-DISKIO-MIB), which net-snmp serves from /proc/diskstats: per device, the number of reads and writes, the bytes read and written, and the microseconds spent busy.

Everything comes from counters, over the interval between two runs: on the first run there is no previous reading, and nothing is returned for that disk; the values come on the next run.
A partition is left out by the code, since its input/output is already counted on its disk: a device whose name is another device's name followed by digits (sda1, vdb2) or by p and digits (nvme0n1p1, mmcblk0p2).
Loop, in-memory and optical devices are left out by the default selection: everything except loop*, ram*, zram*, sr* and fd*. Device-mapper volumes (dm-*, LVM) and software RAID (md*) are kept. The selection can be changed on the probe.
Busy time comes from diskIOBusyTime, which recent agents serve: without it, the busy share is not returned.
A host whose agent serves no disk table stops the collector as misconfigured.

Prerequisites

snmpd (net-snmp) running on the host, built with the ucd-snmp/diskio module (the case in the main distributions), listening on an address the poller reaches (UDP 161), with the firewall open to the poller only.
The view granted to the poller includes diskIOTable (.1.3.6.1.4.1.2021.13.15).
Read-only access, SNMPv2c or SNMPv3, set on the host or one of its folders (protocol:snmp:* variables).

# /etc/snmp/snmpd.conf
agentaddress udp:161
view   banshee included .1.3.6.1.4.1.2021.13.15
rocommunity <community> 192.0.2.10 -V banshee
# or: rouser <user> priv -V banshee

Outputs

Name	Label	Unit	Per
`read-ops`	Reads	op/s	disk
`write-ops`	Writes	op/s	disk
`ops`	Operations	op/s	disk
`read-rate`	Read throughput	MiB/s	disk
`write-rate`	Write throughput	MiB/s	disk
`busy`	Busy time	%	disk

Watches

Watch	Kind	Defaults	Checked
Disk saturated	Above a threshold, on `busy`	warning 90 %, for 5 min	yes
Too many operations	Above a threshold, on `ops`	no threshold	no

Limits

The operation counts are 32-bit counters: on a busy disk they wrap, and the run where they do returns nothing for that disk.
Neither latency nor queue length is served by this MIB.
A disk that is always busy is not always saturated: an SSD or an array serves many requests at once, and reaches 100 % well before its limit.