A link-level pause mechanism that stops traffic on one priority class instead of dropping it — and the source of most RoCE pain.
What PFC Actually Sends
Ordinary Ethernet flow control halts a whole link. PFC halts one of eight priority classes: a single frame carries a pause duration for each class, so RDMA traffic can be stopped while management and storage keep moving. Durations are expressed in quanta, one quantum being the time to put 512 bits on the wire — so the same numeric value means a shorter pause at a higher port speed.
A duration of zero is the resume signal. The point of the whole mechanism is to convert loss into delay, because RDMA survives delay and does not survive loss.
The Side Effects Are the Point
Pausing an upstream port stops every flow in that class on that port, including flows destined somewhere entirely uncongested. That is head-of-line blocking, and it is not a bug — the mechanism has no way to distinguish them.
Worse, the pause propagates. The upstream switch now fills its own buffer, pauses its upstream, and one hot link can throttle a large region of the fabric within milliseconds. In a topology with a cycle, PFC can deadlock outright, each switch waiting on the next, which is why every serious deployment runs a PFC watchdog that force-drains a queue paused for too long.
Why Headroom Depends on Cable Length
The pause frame takes time to reach the sender, and the sender keeps transmitting until it arrives. Every one of those frames must fit in reserved buffer at the receiver, or the pause achieves nothing.
That reserve scales with link rate and with distance, since light covers about 200 metres per microsecond in fibre and a long cross-hall run holds correspondingly more data at the instant of the pause. Vendor defaults assume short links. Extending a fabric across a hall without revisiting headroom is a well-worn way to lose losslessness on precisely the longest, most important links.
What Goes Wrong
The reliable indicator of an unhealthy fabric is a PFC pause counter that climbs. Not a high one — any nonzero rate that persists means ECN is not doing its job and the emergency brake is being used as a throttle.
The other common cause is not congestion at all. A marginal optical link burning FEC correction budget reduces effective throughput on one lane, which fills a buffer upstream and fires PFC exactly as real congestion would. Check optical receive power and error counters before touching thresholds; the RoCE PFC and ECN tuning guide and the link-down diagnostic guide both cover the sequence.
PFC reference
| Standard | IEEE 802.1Qbb |
|---|---|
| Granularity | eight priority classes, paused independently |
| Pause quantum | the time to transmit 512 bits at the port rate |
| Frame content | a pause duration per class, in quanta |
| Resume | a quantum value of zero |
| Predecessor | 802.3x PAUSE, which stopped the entire link |
| Required reserve | headroom buffer for in-flight frames |
| Usual safeguard | PFC watchdog — a vendor feature, not part of 802.1Qbb |
| Health metric | pause counters should stay flat, not merely low |
Related Terms
ECN, lossless Ethernet, RoCE, UEC, FEC.