Packet loss, latency and jitter: three numbers, three different jobs

Voice, video, storage replication, and industrial cyclic I/O do not fail the same way. A few lost packets can freeze a motion axis; the same loss is invisible to a file copy that retries. A stable 8 ms one-way delay may be acceptable for a historian and fatal for a drive telegram that expected 1 ms. Jitter — variation of delay — breaks playout buffers and breaks isochronous control even when average latency looks fine. Network performance work maps those three quantities onto the process that cares.

How the investigation is actually run

Name the application before you name the switch. If it is cyclic I/O, write the period and the payload size. If it is voice, write the codec and whether silence suppression is on. If it is storage, write whether the flow is latency-sensitive or merely large. Then pick two observation points and stay there long enough for the process to fail, or admit that you only saw a healthy hour.

Queueing discipline is a configuration fact: look at it. A remark on a packet is not a guarantee that the hardware honours it. Wireless hops need their own paragraph; they inject jitter that a wired core cannot explain. Electrical retries belong in the EMI file when CRC and drive starts share a clock.

The written conclusion should be usable by both IT and operations: this path loses frames under that backup; this path delays because of that trombone; this path jitters because of mixed queues. Next action is one of those three, not “upgrade the network” as a mood.

Symptoms

Users say “slow.” Cameras stutter. PLC scans complete but the HMI paints late. Backups run long at the same hour as a backup window on a shared core. OT reports “communication timeout” while IT’s ping to the same subnet looks clean because ping is neither the protocol nor the priority queue. See OT/IT and industrial Ethernet.

Causes

Loss: errors on a link (see intermittent faults and active diagnostics), buffer overflow, or a policer. Latency: serialization on a slow hop, extra routing, firewall inspection, or a path that trombones through a data centre. Jitter: mixed queues, Wi-Fi contention, garbage collection on a virtual host, or electrical noise that causes retries — retries look like jitter and loss together; EMI and SNR belong in that file.

Investigation

Measure one-way delay if the process is one-way; round-trip ping averages hide asymmetry. Capture at two points: talker and listener. Mark DSCP/CoS and check whether the switch actually honours it. Correlate with interface errors, CPU, and electrical events. In a plant, confirm cycle time and telegram size before you declare the LAN “too slow.”

Conclusions name which of the three numbers moved, on which path, under which load. They do not quote a marketing SLA. Next step may be queue design, a dedicated VLAN, a physical repair, or an electrical investigation — written as such.