NIC tuning (ethtool)

On a dedicated server with heavy traffic (CDN, streaming, VPN terminator), default NIC settings may limit throughput or increase latency. ethtool inspects and changes link and hardware offload settings.

Prerequisites

• Linux dedicated server, identified interface (ip link)
• root access
• Backup current settings before changes
• Understanding: some offloads break bridging/VPN if misconfigured

Identification

ip -br link
ethtool eth0
ethtool -i eth0

Note: driver, firmware, Speed, Duplex, Link detected.

Link diagnostics

ethtool eth0 | grep -E 'Speed|Duplex|Auto|Link'
ethtool -S eth0 | less

Error counters:

ethtool -S eth0 | grep -iE 'err|drop|fifo'

CRC errors → cable, SFP, switch port. Discuss with HolyCloud if hardware is leased.

Ring buffers (rx/tx)

Increasing rings reduces drops under bursts:

ethtool -g eth0
ethtool -G eth0 rx 4096 tx 4096

Max values per card:

ethtool -g eth0 | head

Persist via systemd or @reboot script — ethtool does not always persist alone.

Example unit /etc/systemd/system/nic-tune.service:

[Unit]
Description=NIC ethtool tuning
After=network.target

[Service]
Type=oneshot
ExecStart=/usr/sbin/ethtool -G eth0 rx 4096 tx 4096
RemainAfterExit=yes

[Install]
WantedBy=multi-user.target

sudo systemctl enable --now nic-tune.service

Offloads: enable or disable?

List:

ethtool -k eth0

| Offload | Role | Disable if… |

|---------|------|----------------|

| tcp-segmentation-offload (TSO) | Outbound TCP segmentation | Fine capture, some tunnels |

| generic-receive-offload (GRO) | Receive aggregation | Specific Open vSwitch bridge |

| lro | Large Receive Offload | Rare asymmetric routing |

Classic 10 GbE web server:

# often leave offloads ON
ethtool -K eth0 tso on gso on gro on

VPN bridge / transparent bridge:

ethtool -K eth0 gro off lro off

Interrupt coalescing

Reduces IRQ under load:

ethtool -c eth0
sudo ethtool -C eth0 rx-usecs 50 tx-usecs 50

Too aggressive → higher latency; measure with mpstat and softnet_stat.

IRQ and RPS/XPS

IRQ affinity (multi-queue):

grep eth0 /proc/interrupts
# set_irq_affinity scripts per Intel/Mellanox vendor

RPS (Receive Packet Steering) if a single queue saturates:

echo 4096 | sudo tee /proc/sys/net/core/rps_sock_flow_entries

Kernel documentation: spread load across CPUs.

Flow control (pause frames)

ethtool -a eth0

In datacenters, often off on both sides to avoid cascade pauses — validate with the operator.

sudo ethtool -A eth0 rx off tx off

Throughput test

iperf3 -s
# external client
iperf3 -c IP_PUBLIQUE -P 8 -t 30

Compare before/after tuning; watch ethtool -S drops.

Multi-gigabit NIC (e.g. Intel X710)

Load correct firmware, use intel-i40e / mlx5 repos per card. HolyCloud lists the chipset on the server sheet.

dmesg | grep -i eth0
lspci | grep -i ethernet

Rollback

sudo ethtool -G eth0 rx 512 tx 512
sudo ethtool -K eth0 gro on gso on tso on

Reboot if the driver stays unstable.

Troubleshooting

| Symptom | Hint |

|----------|-------|

| Link flapping | SFP cable, auto-negotiation, ethtool -r eth0 reset |

| Throughput capped at 100M | Negotiation, cat5 cable |

| Loss after ethtool -G | Value too high — revert to allowed max |

Need help?

Provide ethtool eth0, ethtool -S eth0, card model (lspci -v), and traffic type for HolyCloud switch port advice.