Network Capacity Reporting: Uplink Utilization, 95th Percentile and Trends

When does an uplink need to be upgraded? Decide by evaluating the business-hours P95, short traffic peaks and packets dropped in the queue together.
Technical summary
Technology
SNMP, IF-MIB (RFC 2863), ifHCInOctets, 95th percentile, LACP, NetFlow
Topic
Producing a LAN capacity report from monitoring data
Level
Intermediate
Prerequisites
  • SNMP-based port monitoring infrastructure
  • At least a few months of historical data
In this article9 sections

The monitoring system produces hundreds of graphs from every port, but most organizations have no report that answers the question “when should we upgrade the uplink?” There are graphs, but no decision. Below we explain how to produce a meaningful capacity report from existing LAN monitoring data: the average vs. peak fallacy, the 95th percentile (P95) calculation, the limits of the sampling interval, interpreting discard and error counters, tracking oversubscription, trend projection and a monthly report template.

What question should a capacity report answer?

The value of a report depends on the decision it drives. The goal isn’t a collection of graphs but a document that gives short, evidence-backed answers to these questions:

  • Which uplinks and server ports are full enough during business hours to affect the user experience?
  • Is the utilization turning into real congestion (packet drops), or is it just high usage?
  • At the current growth rate, which link will reach the threshold, and when?
  • Given procurement and installation time, which investment decision should be made this quarter?

The scope is existing LAN and data center ports. Application-based sizing of a new internet or WAN line requires a different calculation; for that, see the internet line capacity planning article.

How utilization is calculated: counters and formula

Port utilization is calculated from the difference between two samples of byte counters read via SNMP. RFC 2863 (IF-MIB) defines two counter families: the 32-bit ifInOctets/ifOutOctets and the 64-bit ifHCInOctets/ifHCOutOctets. For speed, ifHighSpeed (in Mbps) is used.

utilization (%) = (Δ ifHCInOctets × 8 × 100) / (Δt [s] × ifHighSpeed × 1,000,000)

# Reading a single counter on Linux (with SNMPv3)
snmpget -v3 -l authPriv -u monitoring -a SHA -A '<auth-password>' -x AES -X '<priv-password>' \
  10.10.99.2 IF-MIB::ifHCInOctets.10101 IF-MIB::ifHighSpeed.10101

32-bit counters wrap around very quickly at high speed. The upper limit of 4,294,967,296 bytes equals about 34.4 gigabits; at a 1 Gbps line rate the counter overflows in about 34 seconds, and at 10 Gbps in about 3.4 seconds. With 5-minute polling, a value calculated with a 32-bit counter is meaningless. This can be behind the complaint “the 10G uplink always looks empty.” 64-bit counters can’t be read with SNMPv1; v2c or v3 is required. For secure SNMP configuration, see the SNMPv3 configuration guide.

Average, peak and P95: a worked calculation

A daily or monthly average almost always misleads capacity decisions; idle hours pull the average down. Worked example: a 10 Gbps uplink carries 6 Gbps for 8 business hours and 0.5 Gbps in the remaining 16 hours. The daily average is (8 × 6 + 16 × 0.5) / 24 = 56 / 24 ≈ 2.33 Gbps. If the report says “average about 2.3 Gbps,” it looks like there’s no problem. Yet users work with a link carrying 6 Gbps for a third of the day.

A single maximum value misleads in the opposite direction: a momentary peak during the night backup or a software rollout doesn’t represent sustained need. P95 is a balanced measure between the two extremes; it gives the value below which 95 percent of the samples in the period fall.

Worked example: in a 30-day month with 5-minute sampling, 30 × 288 = 8,640 samples are produced. 5 percent of these is 432 samples. All samples are sorted from highest to lowest, the top 432 are discarded, and the highest remaining value (the 433rd) is reported as P95. The sampling time of the 432 discarded samples totals 36 hours. This doesn’t mean 36 hours of saturation in one stretch; the samples may be short peaks spread across the month. A P95 calculated from five-minute averages doesn’t, on its own, show millisecond-level traffic bursts.

The sampling interval and the invisibility of short traffic bursts (microbursts)

5-minute sampling gives the average of those 5 minutes. Worked example: if the 5-minute average on a 1 Gbps port shows 30 percent, that means about 90 gigabits of data in 300 seconds. This data may have been spread evenly over 300 seconds; it may also have flowed at line rate for 90 seconds and been zero for the rest. In the second case, users experience a full link for 90 seconds while the graph shows 30 percent.

At a shorter scale, millisecond-level traffic bursts don’t appear in any SNMP sampling. Multiple 10 Gbps servers sending data to a single 10 Gbps port at the same time can fill the switch buffer within milliseconds and drop packets. The only trace on the SNMP side is the discard counter. Lowering the sampling interval to 1 minute on critical uplinks makes minute-level peaks visible but doesn’t show traffic bursts; for that you need the vendor’s buffer and queue statistics or flow/telemetry data. The graph doesn’t show everything.

Reading discard, error and queue counters

The utilization rate alone doesn’t prove congestion. If a port that is 90 percent full isn’t dropping packets, the user experience may not suffer; a port averaging 40 percent may be dropping because of traffic bursts. That’s why a capacity report always presents utilization data together with these counters:

Counter (IF-MIB) What it tells you Likely cause
ifOutDiscards Packets dropped because there was no room left in the output queue Congestion, traffic bursts, QoS policy
ifInDiscards Packets received but dropped before they could be processed Input buffer, resource limit
ifInErrors Frames received with errors (CRC, etc.) Cable, optics, duplex mismatch
ifOutErrors Frames that couldn’t be transmitted Physical layer, driver

Error counters are a physical layer problem, not a capacity one. A port showing rising ifInErrors isn’t upgraded; its cable or optic is checked. A mistake you may encounter is adding a second link as the solution to an uplink with rising CRC errors, when cleaning a dirty fiber connector would have been enough. On Cisco devices, the same information is seen in “Total output drops” in the show interfaces output and with show interfaces counters errors. Counters are reported not as absolute values but as the rate of increase within the period; also remember that the clear counters command can break the calculation in the monitoring system.

The oversubscription ratio, in an enterprise network infrastructure, is the ratio of the total capacity of access ports to uplink capacity. Worked example: on a switch with 48 × 1 Gbps access ports and 2 × 10 Gbps uplinks, the ratio is 48 / 20 = 2.4:1. The ratio alone isn’t good or bad; because most user traffic doesn’t peak at the same time, a certain oversubscription at the access layer is normal. The decision is driven not by the ratio itself but by the uplink’s actual P95 and discard counters.

If uplinks are a port-channel, each member link is monitored separately alongside the total. Because LACP load balancing uses flow-based hashing, traffic may not be distributed evenly across members; one member may be full while the total shows 45 percent. The total graph hides this. For choosing the hash method, see the LACP and EtherChannel configuration article. To see which flows are filling the uplink, NetFlow, sFlow and IPFIX data is used; before an upgrade decision, it may turn out that part of the traffic is taking the wrong path or coming from an unplanned backup job.

Setting thresholds and trend projection

Fixed rules such as “upgrade when it exceeds 70 percent” don’t fit every link. The threshold is set per link with these inputs: the latency sensitivity of the traffic carried (voice and video, or backup), the utilization level at which discard counters start to rise, whether the remaining link can carry the full load when one link goes down in a redundant design, and the procurement–installation time. The last two items are often skipped. If each link in a redundant 2 × 10 Gbps uplink carries 6 Gbps, when one goes down, the 12 Gbps load doesn’t fit into a single 10 Gbps. Redundancy is lost at that moment.

For trends, monthly P95 values are kept for at least 6–12 months and a simple linear trend is derived. Worked example: suppose a core uplink’s business-hours P95 is 4.2 Gbps this month, the average increase over the last 12 months is 0.15 Gbps per month, and the threshold set for this link is 7 Gbps. The time remaining to the threshold is (7 − 4.2) / 0.15 ≈ 18.7 months. If procurement and installation take 4 months, the decision must be made within about 14 months at the latest. Linear projection doesn’t catch jumps such as moving a new location or migrating an application to the cloud; such planned changes are added to the report as a separate line.

Monthly capacity report template

The report has two layers: a summary page for management and a detailed table for the technical team. In the detailed table, each critical link being monitored is a single row:

Field Content
Connectivity Device, port, far end, capacity
Use Business-hours P95, monthly peak (inbound/outbound separately)
Quality Discard and error increase, incident note if any
Trend Trend over the last 6–12 months, estimated time remaining to the threshold
Status and action Normal / under watch / action required, recommended step

The summary page contains only the links in “action required” and “under watch” status, the recommended step and the decision date. The monitoring platform that produces the report must support percentile calculation and long-term data retention. Some systems store old data by thinning it out (averaging); past peaks are erased that way. When choosing a platform, evaluate this criterion together with the Zabbix, LibreNMS and PRTG comparison.

The relevant pages are our network maintenance, monitoring and periodic checks service for tying capacity reporting into the regular maintenance cycle, and our network performance analysis and troubleshooting service for diagnosing performance problems caused by discards and traffic bursts.

Frequently asked questions

Why is P95 a better capacity measure than the average?

The average hides busy periods because of idle hours; a single maximum value is affected by temporary peaks. P95 gives the value below which 95 percent of the samples in the period fall; it leaves out short-lived peaks while reflecting sustained load.

Utilization on the graph is low but users experience slowness; why might that be?

A 5-minute average hides peaks at the minute and millisecond level. If the ifOutDiscards counter on the same port is increasing, packets may be dropping due to traffic bursts. One port-channel member being full because of unbalanced load gives the same symptom.

Why are 64-bit counters essential for capacity measurement with SNMP?

A 32-bit byte counter wraps around in about 34 seconds at 1 Gbps and about 3.4 seconds at 10 Gbps. With polling intervals lasting minutes, the calculated value becomes meaningless. The 64-bit ifHCInOctets and ifHCOutOctets counters eliminate this problem.

There is no single value that applies to everyone. The threshold is set per link by looking at the latency sensitivity of the traffic carried, the utilization level at which discard counters start to rise, the remaining capacity when one of the redundant links goes down, and the procurement time.

Support

Contact us for support or a discovery session on this topic.

Contact form
[email protected]

Eretek Network

We build switching, routing, cabling and wireless networks with VLAN segmentation and redundant uplinks, and provide regular maintenance and fault support.

Knowledge Center

Other articles