The network works but training crawls
A network link can remain “up” while errors affect a training job. Start with fresh evidence from the affected path.
Is the counter increasing now
- Start
1,000 errors
- Ten minutes later
1,600 errors
- Difference
600 errors / 600 seconds = 1 per second
Read the change in counters
An old error total does not show whether errors are occurring now. Compare readings over a known interval and account for resets.
FEC, forward error correction, can correct certain transmission errors. Corrected and uncorrectable errors mean different things; use the relevant device guidance.
Map the job to the link
Identify workers, ports, both endpoints, negotiated speed, and recent link transitions. Compare the time of new errors with the slowdown.
Keep other explanations open: congestion, placement, software, storage, or a late worker can also delay training.
Intrusive tests and cable replacements require coordination and approved maintenance.
Try it yourself
A fictional counter rises from 1,000 to 1,600 in ten minutes without a reset. What is the average increase?
Check your answer
One error per second. Severity still depends on the counter, traffic, vendor limits, and impact. A total staying at 1,000 during a later slow job does not establish new errors in that interval.