Run ip -s link show IFACE and watch whether the RX/TX errors and dropped counters increase. Errors usually point at a physical problem (cable, NIC, duplex mismatch); drops usually mean the host could not process packets fast enough — a full ring buffer, a saturated link or a CPU busy with softirqs.
Symptoms
- Increasing dropped or errors counters in ip -s link
- Packet loss or retransmits under load
- Intermittent slowness that disappears at low traffic
How to diagnose it
Check interface counters
Run twice a minute apart; only increasing counters matter.
ip -s link show eth0See detailed NIC statistics
Look for rx_missed, rx_no_buffer, fifo or crc counters.
sudo ethtool -S eth0 | grep -iE "drop|err|miss|fifo|crc" | grep -v ": 0"Check link speed and duplex
Half duplex or a lower speed than expected suggests a negotiation or cable problem.
sudo ethtool eth0 | grep -E "Speed|Duplex"Check TCP retransmissions
Rising retransmits confirm loss is affecting connections.
nstat -az | grep -i retrans
Common causes and fixes
Ring buffer too small for bursts
Increase the RX ring to the hardware maximum shown by ethtool -g.
sudo ethtool -g eth0
sudo ethtool -G eth0 rx 4096Link saturated
Compare throughput with the link speed. Shape or move bulk traffic (backups), or upgrade the link.
Bad cable, port or duplex mismatch
CRC or frame errors point at hardware. Replace the cable, try another switch port, and make sure both ends auto-negotiate.
Firewall or conntrack table full
Kernel log shows “nf_conntrack: table full, dropping packet”. Raise the limit or reduce tracked connections.
sudo dmesg -T | grep conntrack
sudo sysctl -w net.netfilter.nf_conntrack_max=262144Prevent it happening again
- Statance records packets per second, errors and drops per interface every minute so you can see when they started.
- Add uptime checks from more than one region to tell host network problems from wider outages.
Set it up: Linux monitoring · Hetzner monitoring.
FAQ
What is the difference between dropped and errors?
Errors are malformed or corrupted frames, usually hardware. Drops are valid packets the host discarded, usually because buffers were full.
Are a few dropped packets normal?
A small, static count is common. Counters that keep rising during normal traffic need investigation.
Related guides
Get alerted before this becomes an outage.
Install the agent in one line. Statance watches CPU, memory, disk and services, and its AI assistant helps find the cause.