Home › Topics › Capacity Planning From Monitoring History › Capacity Beyond Bandwidth

Capacity Beyond Bandwidth: Disk, Ports, Addresses, Licences, Power


Ask most administrators what "capacity" means and they will describe a WAN link. That is fair; the link has the best graphs and the loudest users. But outages that come from something being full are mostly not bandwidth outages. They are a volume that hit one hundred percent or a switch with no free port on the day the printer arrived. They are a DHCP scope that ran dry during a visitor event or a closet that got too warm in the summer. Each is capacity, each leaves a trail in monitoring, and each has a lead time.


This guide walks through the resources that fill up besides bandwidth, one section each. For every resource you will learn what full looks like, what to measure, and which counter or check gives you the number. You will learn the warning that arrives first and the surprise factor: how likely it is to catch you out. At the end there is a table template and a story about a closet that ran out of ports two weeks after an audit counted twelve spare.


In this guide

  • The same five questions for every resource
  • Storage: volumes, backup targets, and the monitoring database
  • Memory and CPU: the monitoring server and the key hosts
  • Switch ports and the PoE budget
  • IP addresses and DHCP scopes
  • Licences and device counts
  • Power, rack space, and cooling
  • The per-resource table template
  • The ones that surprise people
  • Checklist
  • Where to go next

The same five questions for every resource


Every resource in this guide is handled with the same five questions. Once you have the habit you can apply it to anything that fills up. What does full look like: a hard wall or a slow degradation? What do I measure, in which units, against which limit? Where does the number come from: which counter, check, or command? What warning arrives first? And what is the surprise factor: how often does this one catch people who were paying attention? The answers differ, and the differences are the whole subject.


The method for turning numbers into a date is the same for all of them and is not repeated here. The forecasting guide covers the trend, the lead time guide covers the trigger, and the basics guide covers the register they both feed. Bandwidth itself is deliberately absent: the bandwidth utilisation monitoring pillar already covers it, including the upgrade-or-shape decision for a hot link.


Storage: volumes, backup targets, and the monitoring database


What full looks like: a hard wall. When a volume reaches one hundred percent, whatever writes to it stops, and it stops immediately. Databases refuse transactions, log files stop recording, updates fail, and Windows becomes inventive. There is no queue, no graceful slowdown; the last byte is the last byte.


What to measure: used bytes against total bytes, per volume, as a daily maximum. Percent is convenient for thresholds, but forecast in bytes. That is because one percent of a two-terabyte volume and of a twenty-terabyte volume are very different amounts of headroom.


Where it comes from: on Windows hosts, WMI or an agent gives the numbers per logical disk. On anything that speaks SNMP, the host resources storage table lists each volume with its size and used count in allocation units. A walk of it looks like this (the names are the standard ones from the host resources MIB):


snmpwalk -v2c -c public fs01.example.com hrStorageDescr
snmpwalk -v2c -c public fs01.example.com hrStorageSize
snmpwalk -v2c -c public fs01.example.com hrStorageUsed
snmpwalk -v2c -c public fs01.example.com hrStorageAllocationUnits

used bytes = hrStorageUsed x hrStorageAllocationUnits
size bytes = hrStorageSize x hrStorageAllocationUnits

The warning that arrives first: the backup job that starts taking longer, because the backup target is where growth shows up twice. Every byte added to a server is added again to the copy of it. And backup targets fill faster than the volumes they protect while attracting less attention. Watch the backup target like any other volume, and the job duration as a second signal.


The surprise factor: low for server volumes, high for the monitoring database. The monitoring system writes every sample into its own store, and adding fifty devices adds their samples forever after. If that store lives on a volume nobody graphs, it fills quietly and takes the history with it. That is an unkind way to lose a capacity plan. Give the monitoring database its own row in the register. That applies whether it is SQLite on a small install or MySQL/MariaDB or SQL Server on a larger one. Check the guide on database backend choices for how each grows.


Memory and CPU: the monitoring server and the key hosts


What full looks like: slow, then slower, then stopped. Neither CPU nor memory has a wall in the way a disk does. A busy processor makes everything wait. A full memory pushes the operating system into swapping, which is a way of turning memory shortage into disk traffic and disk traffic into misery. "Full" is the point at which the busy hour makes the service unusable, and that point is a judgement, not a number.


What to measure: the busy-hour peak of CPU percent and of memory in use, per host, per day. Never the daily average. A server averaging thirty percent CPU with an hour a day at ninety-eight has a capacity problem for one hour a day. That is the hour the users are there. Memory is measured in bytes in use against bytes installed. On hosts with a page file, the rate of paging is the honest indicator of pressure.


Where it comes from: WMI on Windows, SNMP or SSH on almost everything else. The monitoring server deserves special attention, because its CPU and memory scale with the number of checks and the polling interval. The guide on polling load budget explains how to work out what each new device costs. The guide on monitoring performance tuning covers what to do when the monitoring server itself is the one filling up.


The warning that arrives first: the peak spreading. The busy hour becomes ninety minutes, then two hours. On a year-long graph of daily maxima the line flattens against a ceiling rather than climbing, because the peak cannot pass one hundred. The tell is the width, not the height. Look at the hours per day above eighty percent, and forecast that instead.


The surprise factor: medium. People expect CPU and memory to grow but forecast them from averages, which hides the peak until it is the whole afternoon.


Switch ports and the PoE budget


What full looks like: a technician standing in a closet holding a cable with nowhere to put it. Ports are slots, so full is exact: forty-eight ports, forty-eight cables, and the new access point waits. The PoE budget is the second, subtler limit. Power over Ethernet delivers electricity to phones, cameras, and access points through the data cable. The switch has a fixed number of watts to share. A switch can have free ports and no free watts, and the device you plug into a free port simply does not power up.


What to measure: for ports, the count of ports in use against the count that exist, per switch. For PoE, the watts currently drawn against the watts the power supply can deliver. Both are read from the switch over SNMP: the interface table tells you each port's operational state. And the power-over-Ethernet table tells you the supply's rated and consumed power.


snmpwalk -v2c -c public switch01.example.com ifOperStatus
   (count of "up" = ports in use today)
snmpwalk -v2c -c public switch01.example.com pethMainPsePower
snmpwalk -v2c -c public switch01.example.com pethMainPseConsumptionPower

free ports = total ports - ports up
PoE headroom (W) = pethMainPsePower - pethMainPseConsumptionPower

Where the trap is: "up" is not the same as "in use". A port can be patched to a wall socket with nothing plugged in, and it shows down while being spoken for. Count a port as free only if it has been down for a month and is not patched. The guide on PoE budgets covers the power side, including why a switch's rated PoE is often less than the sum its ports could draw.


The warning that arrives first: the last-free-port conversations. "Can I borrow the port the old printer was on?" is the sound of a switch with two ports left. On the PoE side, the first sign is a device that powers up on one port and not another. That gets logged as a faulty port and is really a budget.


The surprise factor: high. Ports are the resource with the longest gap between how easy they are to count and how rarely anyone counts them.


IP addresses and DHCP scopes


What full looks like: a network that works perfectly for everyone who already has an address and not at all for anyone new. A DHCP scope is the pool of addresses a server hands out on a subnet. When the pool is empty the next device to ask gets nothing. Existing devices keep working until their leases expire, so the outage arrives one laptop at a time. The tickets say "my machine works, his does not."


What to measure: addresses leased against addresses in the pool, per scope, as a daily maximum. Also count the subnet's fixed addresses (servers, printers, reservations) against the subnet's total. That is because a subnet can be full of static assignments while the scope looks half empty. Windows DHCP servers report the figures directly:


Get-DhcpServerv4ScopeStatistics |
  Select-Object ScopeId, InUse, Free, PercentageInUse

Where it comes from: the DHCP server, by the command above or by a monitoring check that asks the same question. The guide on DHCP scope monitoring covers setting that up, and scope exhaustion covers the failure itself. This guide only asks you to forecast the date.


The warning that arrives first: a scope that fills on a Tuesday afternoon and empties overnight. The daily maximum creeps up while the average looks fine, because leases from visitors and phones are short-lived. Forecast on the peak. Lease time matters too: a scope with long leases and a busy lobby can fill at lunchtime.


The surprise factor: high, because nothing about a network looks different at ninety-five percent scope use. The graph is the only warning, and the graph is only there if someone is collecting it.


Licences and device counts


What full looks like: the product stops doing something. Depending on how the licence is counted, that could be refusing to add the new switch or refusing to poll the new interfaces. Or it could be carrying on while quietly recording that you are out of compliance. Full is defined by the licence agreement. The first task is to read it and write down, in one line, what happens when the count is exceeded.


What to measure: licensed units consumed against licensed units purchased, in whatever unit the product counts: devices, sensors, interfaces, users, cores. The trend comes from the rate you add devices, which the monitoring system's own discovery history shows you. Most monitoring tools show their own licence use in the console; copy it into the register monthly if there is no counter.


The warning that arrives first: the discovery run that finds more than you expected. A merger, a new site, a batch of access points, or a switch replaced with a bigger one all add units in a single step. A step is exactly what the ruler cannot see coming. The licensing models and hidden costs pillar covers the ways counts creep upward and where the tier boundaries fall.


The surprise factor: medium. The number is easy to find and nobody looks at it until the day it refuses something.


Power, rack space, and cooling


What full looks like: three different things. Power full is a UPS or a circuit at its rated load. It trips or, worse, drains the battery in minutes during the first outage instead of the half hour you planned. Rack full is a physical limit measured in rack units, the standard height increments equipment is built to. Cooling full is a room whose temperature climbs every afternoon and no longer comes back down overnight. That is the one of the three that shortens the life of everything else in the room.


What to measure: for power, measure the UPS output load as a percentage of its rating, as a daily maximum. Read it over SNMP from the standard UPS table or the management card. For rack space, free units per rack, counted by hand on every install. For cooling, the room or inlet temperature as a daily maximum and minimum, because a climbing minimum means the room is no longer recovering. The guides on UPS monitoring and temperature and humidity cover collecting these.


The warning that arrives first: for power, a UPS runtime estimate that has been shrinking for months, because every added device shortens the same battery. For cooling, the overnight minimum. Suppose a room used to cool to eighteen degrees by morning and now only reaches twenty-two. It is telling you the air conditioning has stopped keeping up, months before the afternoon peak becomes a problem.


The surprise factor: high for cooling, medium for power, low for rack space. Nobody forecasts cooling because the number does not grow in the way a disk grows. It grows with every device installed, and the effect only shows in summer. (The equipment notices before the people do, and it complains by failing.)


The per-resource table template


Here is the template for the "what to measure" side of the register: one row per resource kind, filled in with the answers from this guide. Keep it next to the register so whoever updates the numbers knows where each one comes from. Replace the source column with the specific counter or check in your own monitoring system.


ResourceMeasureSource"Full" definitionFirst warning
Server volumeUsed bytes vs total, daily maxWMI, agent, or SNMP host resources table100% used; writes failBackup duration rising
Backup targetUsed bytes vs total, daily maxSame as volume, on the backup host100% used; jobs failRetention silently shortened
Monitoring databaseDatabase file or tablespace size vs volumeDatabase status, volume counterVolume full; collection stopsGrowth step after adding devices
CPU on key hostBusy-hour peak percent; hours/day above 80%WMI, SNMP, or SSHPeak makes service unusablePeak widening, not rising
Memory on key hostBytes in use vs installed; paging rateWMI, SNMP, or SSHSustained pagingPaging on busy days
Switch portsPorts up vs total; unpatched portsSNMP ifOperStatus; patch recordsNo unpatched port leftRequests to borrow ports
PoE budgetWatts consumed vs ratedSNMP PoE main PSE tableNext device will not powerDevice powers on one port only
DHCP scopeLeases in use vs pool, daily maxDHCP server statisticsNo free lease at peakPeak rising while average flat
Subnet static spaceFixed addresses vs subnet sizeIPAM or address spreadsheetNo free static addressReservations placed in scope range
Licence countUnits consumed vs purchasedProduct console; discovery historyAs the agreement defines itDiscovery finds a step
UPS loadOutput percent load, daily maxSNMP UPS table or management cardRated load; runtime too shortRuntime estimate shrinking
Rack spaceFree rack units per rackHand count on each installNo contiguous free unitsEquipment on the floor
CoolingInlet temperature, daily max and minEnvironmental sensor or device sensorPeak above equipment ratingOvernight minimum climbing

Tip: two rows in that table, rack space and static addresses, come from hand-kept records rather than a counter. Give them a register row anyway and update them on the monthly day. A resource without a graph still has a trend; it is just kept in a slower kind of database.


The ones that surprise people


Three resources on the list have a surprise factor out of all proportion to how hard they are to count: switch ports, DHCP scopes, and cooling. They share a cause: none produces a graph unless someone collects one. None degrades before it fails, and all are checked by audit rather than by monitoring. An audit is a snapshot, and a snapshot of a slot resource is out of date the moment the next device arrives.


Harbourview Clinic had an audit that counted twelve spare ports in the second-floor closet, and the audit was correct on the day it was done. Over the following fortnight, a new imaging cart arrived, along with four replacement phones that needed their own drops. There were also two conference-room displays and a wireless refresh that put access points where none had been. Together, these consumed eleven of the spare ports. The twelfth went to a contractor's laptop, temporarily. On the day the clinic's new reception printer arrived, the technician stood in the closet holding a cable. The switch that would have solved it had a five-week lead time. The audit was filed under "complete". (It was, briefly.)


The remedy is a counter, not a better audit. Collect a daily port count from the switch, a daily scope count from the DHCP server, and an hourly temperature from a sensor. Those turn three surprise resources into three ordinary register rows. Then the ruler works on them like anything else.


Checklist


  • List every volume that would stop a service if it filled, including backup targets and the monitoring database's own volume.
  • Collect CPU and memory as busy-hour peaks, plus hours per day above eighty percent for hosts near their ceiling.
  • Collect ports up per switch daily, and keep a patch record so down and free are not confused.
  • Collect PoE consumed against rated per switch.
  • Collect leases in use per scope daily, and note the lease time next to each scope.
  • Record licence units consumed against purchased monthly, and note what the agreement says happens when they are exceeded.
  • Collect UPS output load and runtime estimate; watch runtime shrink as a leading sign.
  • Collect room or inlet temperature and forecast on the overnight minimum as well as the afternoon maximum.
  • Keep rack units and static address space as hand-updated rows, refreshed on the monthly register day.
  • Give every row in this guide a lead time, because switches, licences, and cooling all have one.

Where to go next


Every resource here fits the same shape: a number, a limit, a source, and a first warning. Once each has a counter it is forecast and bought like anything else. The rest of Capacity Planning From Monitoring History covers those shared steps. And the guide on justifying an upgrade with graphs shows how to present any one of these rows to the person who signs the order. Start by adding the three surprise rows to your register this week. Ports, scopes, and cooling have never once announced themselves in advance, and they are not going to start now.


Put this into practice with SysPeep Monitor

Most rows in this guide come from counters a monitoring system already collects. Those include disk over WMI or SNMP, port state and PoE from switches, plus UPS load and temperature from their sensors. SysPeep Monitor auto-discovers devices and interfaces, collects these metrics over SNMP, WMI, and SSH, and keeps the long-term graphs in one web console. It lets you set thresholds on utilisation so a scope or a closet gets a planning alert instead of a surprise. The free trial installs on any Windows machine.

Download Free TrialLearn About SysPeep