The SLA annex of a carrier contract is read once on signing day, then stays in a drawer until the first major outage. SLA evaluation is based on the measurement point, period and method in the contract. Your own measurements can support the fault ticket you open with the carrier. Below we explain how latency, jitter, packet loss and availability are defined on WAN and internet lines, how to measure them with Cisco IP SLA, TWAMP and iperf3, and how to prepare evidence that will be accepted when reporting a fault to the carrier.
Metrics measured in a carrier SLA
When reading the SLA annex, ask four questions for each metric: How is it defined? Between which two points is it measured? Over what period is it averaged? What happens if it’s breached? A typical enterprise line contract contains these items:
| Metric | Point to check in the contract |
|---|---|
| Reachability | Is it calculated monthly or yearly; is planned maintenance excluded? |
| Latency | One-way or round trip; PE–PE or CE–CE? |
| Jitter | By which method, and for which traffic class? |
| Packet loss | Averaging period and measurement packet frequency |
| Time to repair | The moment the clock starts: opening the fault ticket, or detection? |
Converting the availability percentage into minutes makes the expectation clear. A 30-day month is 43,200 minutes; a monthly 99.9 percent commitment is one thousandth of this, i.e. an outage allowance of about 43 minutes. The measurement point is the detail most often overlooked. In MPLS services, carriers mostly give the latency commitment between PE routers in their own backbone; the access circuit and the equipment at the customer site are outside this measurement. This is the first reason your end-to-end measurement and the carrier report come out different. Both may be correct.
Definitions of latency, jitter and packet loss
The standard definitions of these metrics are in the documents of the IETF IPPM working group. They’re the common language when discussing with the carrier:
- One-way delay (RFC 7679): The time between a packet leaving the source and arriving at the destination. The clocks at the two ends must be synchronized; because NTP can leave millisecond-level error, interpret the result carefully on low-latency lines.
- Round-trip delay (RTT) (RFC 2681): The most common measurement because it doesn’t require clock synchronization; but it doesn’t show in which direction the delay occurs.
- Jitter (IPDV) (RFC 3393): The difference between the delays of consecutive packets. In voice and video applications, the jitter buffer absorbs this variation; when the buffer is exceeded, a packet counts as lost even though it arrived late.
- Packet loss (RFC 7680): The ratio of packets that don’t reach the destination within a given time. The distribution matters too: the same average rate affects voice quality little with scattered single losses, but with consecutive losses it causes noticeable gaps in conversation.
As a reference for voice, ITU-T G.114 recommends keeping one-way delay below 150 ms for most applications. This isn’t a carrier commitment; it’s an end-to-end budget; encoding, the jitter buffer and LAN delay also take a share of this budget. The share left for the line itself is smaller.
Active measurement: Cisco IP SLA and TWAMP
Interface counters and flow data tell you how full the line is, not how good it is. Seeing quality requires active measurement that regularly sends synthetic test traffic over the line.
Cisco IP SLA
On Cisco IOS and IOS XE routers, IP SLA offers various operations from ICMP to UDP jitter measurement. The udp-jitter operation gives latency, jitter and per-direction loss; an IP SLA responder must be running at the other end. The example below sends test traffic imitating a voice codec to the router at headquarters and raises an SNMP trap when the latency threshold is exceeded:
! Branch router
ip sla 10
icmp-echo 198.51.100.1 source-interface GigabitEthernet0/0/0
frequency 30
ip sla schedule 10 life forever start-time now
!
ip sla 20
udp-jitter 10.255.0.1 16384 codec g711alaw
tos 184
frequency 60
ip sla schedule 20 life forever start-time now
ip sla reaction-configuration 20 react rtt threshold-value 150 100 threshold-type consecutive 3 action-type trapOnly
!
! Headquarters router (far end)
ip sla responder
!
! Reviewing the results
show ip sla statistics 20 details
show ip sla summary
The tos 184 value corresponds to DSCP EF (46) marking; measurement packets are carried in the same QoS class as real voice traffic. A mistake you may encounter is running the measurement without marking: measuring in the best-effort class on a line with QoS doesn’t reflect the experience of voice traffic. Check the responder side too: if there’s no ip sla responder at the far end, or a firewall in between blocks the test port, the udp-jitter operation fails silently. Always read the return code in the show ip sla statistics output.
TWAMP
TWAMP (RFC 5357) is the standard protocol for two-way measurement between devices from different vendors. Besides the full version made up of control and test sessions, the simpler TWAMP Light mode, where only the reflector side runs, is also common; the newer STAMP (RFC 8762) standardizes a similar lightweight approach. Many carrier CPEs and SD-WAN devices have a TWAMP reflector. If you want to measure with the same method as the carrier, ask about TWAMP support on the handoff device at the contract stage. For protocol details, RFC 5357 is the primary source.
Throughput testing with iperf3
Ping measures availability and ICMP round-trip time. IP SLA udp-jitter and TWAMP measure latency, delay variation and loss. iperf3 measures the throughput achieved at the application level. A high-volume iperf3 test fills the line; on a production line, run it only in a maintenance window or with a rate limit. iperf3 runs as a server at one end and a client at the other:
# At the far end (headquarters or cloud server)
iperf3 -s
# TCP, 4 parallel streams, 30 seconds, upload direction
iperf3 -c 203.0.113.50 -P 4 -t 30
# The same test in the download direction (the server sends)
iperf3 -c 203.0.113.50 -P 4 -t 30 -R
# UDP at a fixed 50 Mbps: jitter and loss report
iperf3 -c 203.0.113.50 -u -b 50M -t 30
# To save the results as JSON
iperf3 -c 203.0.113.50 -P 4 -t 30 -J > test-$(date +%F-%H%M).json
Know the limits when interpreting the result. A single TCP flow may not reach line speed on high-latency lines because of the window size; parallel streams reduce this effect. The test client should be not on the wireless network but wired, and at a point that doesn’t carry the inspection load of the firewall in front of the line. If the firewall’s IPS or SSL inspection capacity interferes with the result, what you’re measuring is the device, not the line. We see this mistake a lot. In a UDP test, setting -b above the line speed produces loss; this loss is the result of overloading, not of the line. How line capacity should be determined from the start is explained in the internet line capacity planning article.
Measurement point and timing
Where and when the measurement is made determines for whom the result will be evidence.
- Measure close to the handoff point: Evidence to be presented to the carrier is strong if the measurement source is right behind the carrier device. The firewall’s outside interface, or a measurement device connected to the handoff switch, is a good choice.
- Keep an end-to-end measurement separately: A second measurement representing user experience lets you separate LAN- and firewall-caused problems in the enterprise network infrastructure from line problems.
- Measure continuously: If the problem appears only at certain times (start of the business day, night backup), spot tests are useless. Low-frequency but continuous active measurement reveals the problem’s time pattern.
- Read it together with load: If a latency increase coincides with your own line utilization, the problem is capacity; if it also appears when the line is empty, look on the carrier side.
To see intermediate points along the path, use a report such as mtr -rwzc 200 203.0.113.50. Loss that appears at only one intermediate hop and not at later hops or the destination is mostly caused by that device processing ICMP addressed to itself at low priority. It isn’t real loss. Going to the carrier with this screenshot is the fastest way to be turned down in the first exchange.
Threshold and alarm definitions
Derive thresholds both from the contract values and from the line’s own normal. Instead of universal constants, take as reference the baseline you derive from a few weeks of measurement. The table below is an example structure; adapt the values to the line type.
| Metric | Warning (example) | Critical (example) |
|---|---|---|
| RTT | Baseline + 20 ms, 3 consecutive measurements | Contract value exceeded, 5 consecutive measurements |
| Jitter | Twice the baseline | A value exceeding the voice buffer |
| Packet loss | Non-zero for 15 minutes | Contract threshold exceeded |
| Reachability | 3 consecutive tests failed | Line down, switched to backup |
An alarm based on a single measurement produces noise; add a condition of consecutive counts or a time window. Define a similar condition for clearing (hysteresis) too, otherwise a line hovering at the boundary produces on-off notifications all night and a week later the team stops reading that alarm. If you’ll keep the measurements in a monitoring system, the options are evaluated in the Zabbix, LibreNMS and PRTG comparison article.
Evidence package for reporting a fault to the carrier
The carrier’s fault desk often questions your side first. Applying with the package below shortens the first rounds of correspondence:
- Line ID and handoff information: Circuit number, handoff interface, IP information.
- Timestamped measurement data: Latency, jitter and loss graphs from NTP-synchronized devices, from before and after the moment the problem started. Write down the time zone.
- Load graph: The line’s utilization over the same period; it shows the problem isn’t caused by your own traffic.
- Physical layer data: Interface error counters (CRC, input errors), optical levels on fiber lines (
show interfaces transceiveron Cisco), link-down records. - Path analysis: mtr or traceroute output in both directions; if possible, also taken from the far end.
- Direct connection test: If needed, a brief test with a laptop connected directly to the handoff point to take the firewall out of the picture; this laptop’s local firewall must be on and its patches current.
In dual-line setups, the same measurements are also inputs to the decision to switch to the backup; this mechanism is explained in the failover with a dual internet line article. Continuous monitoring of line quality and managing carrier processes fall within our internet line management service, and diagnosing existing performance problems within our network performance analysis service.
Frequently asked questions
Does latency measured with ping replace SLA measurement?
Partly. ICMP-based RTT measurement is enough for tracking availability and rough latency; but it gives limited information on jitter and one-way loss, and some devices process ICMP at low priority. On lines carrying voice and video, UDP-based measurement (IP SLA udp-jitter or TWAMP) gives more accurate results.
The carrier doesn’t see a problem in its own measurement; what should I do?
First match up what the two sides are measuring: the carrier often measures between PEs in its own backbone, while you measure end to end. Measure as close to the handoff point as possible; apply with timestamped, continuous data, the line ID and physical interface error counters. Data showing that the problem is beyond the handoff point speeds up acceptance of the fault.
Is it right to run iperf3 tests against open servers on the internet?
The result is also limited by the load on the other server and its internet connection; it isn’t reliable for proving line capacity. Use a server at your other location or in the cloud under your own control, with a faster connection than the line being tested, or ask the carrier for a test server.
What does an availability percentage mean in practice?
On a monthly basis, in a 30-day month, 99.9 percent availability means a total outage allowance of about 43 minutes, 99.95 percent about 22 minutes, and 99.99 percent about 4.3 minutes. Separately check in the contract whether planned maintenance and customer-caused faults are excluded from this calculation.